Skip to content
IAM Insights

Explainers · Glossary

Model Context Protocol

MCP

By Boston Identity · Explained · 7 min read

The short answer

The Model Context Protocol (MCP) is an open standard for connecting AI applications to tools and data. An AI app's client calls MCP servers that offer tools; protected servers accept only OAuth tokens issued for them.

What it is
A protocol
Defined by
MCP project specification, revision 2026-07-28
Who it's for
Teams connecting AI assistants and agents to company tools and data
In our briefings
252By month of the source: 2025-12: 8, 2026-01: 8, 2026-02: 13, 2026-03: 13, 2026-04: 16, 2026-05: 9, 2026-06: 14, 2026-07: 15, 2026-08: 9, 2026-09: 31, 2026-10: 42, 2026-11: 33by month, last 12
01

What is Model Context Protocol?

MCP is an open protocol for connecting applications built on large language models (LLMs) to outside data sources and tools.S1 An AI coding tool, a chat app or a custom AI workflow can use it as one standard way to give a model the context it needs.S1

The application the user works in is the host, and it starts the connections.S1 Inside the host, a client is the connector that talks to an MCP server, a service that offers context and capabilities; the messages between them are JSON-RPC 2.0.S1 A server can offer resources (data), prompts (message templates) and tools, which are functions the model can run.S1

Authorization is optional in MCP.S4 When a server on an HTTP connection uses it, the server acts as an OAuth 2.1 resource server, the service that accepts access tokens, and the client acts as an OAuth 2.1 client on the user's behalf.S4 §Roles The specification does not cover how the authorization server is built.S4 §Roles It does require that server, which signs the user in and issues tokens, to implement OAuth 2.1 and to publish discovery metadata under RFC 8414 or OpenID Connect Discovery.S4 §Overview An implementation that connects over stdio (standard input and output) should not use this flow, and gets its credentials from the environment instead.S4

MCP is not a policy system. It says how a client gets a client ID and an OAuth token and how it presents the token.S4 The decision about who may do what is moving up into identity providers and agent gateways.1

02

How does Model Context Protocol work?

We follow one host application, with its MCP client, as it reaches a protected MCP server over HTTP and gets a token from that server's authorization server. The steps use the current revision, in which every request stands on its own.S2

Figure steps
  1. The client checks the protocol version
  2. The server points to its authorization server
  3. The client registers and the user signs in
  4. The client lists the tools
  5. The model picks a tool and the client calls it
  6. The server asks for more scope when it needs it
Fig. 1

The MCP client checks the version, gets a token for one server from that server's authorization server, then lists and calls tools, and asks for more scope only when the server names what is missing.

Drawn from the MCP specification, revision 2026-07-28

Swipe for the full figure →

  1. The client checks the protocol version

    Every request names the protocol version it uses, in its _meta field and, on HTTP, in the MCP-Protocol-Version header.S2 A server that does not support that version must answer with an error that lists the versions it does support, and the client should retry with one they share.S2 Servers must offer a server/discover call, which a client may send first to learn those versions.S2 Revisions up to 2025-11-25 opened with an initialize handshake; the current revision has none.S2

  2. The server points to its authorization server

    A protected server must answer a request with a missing, invalid or expired token with HTTP 401.S4 Its WWW-Authenticate header should name the scopes the request needs.S4 The server must publish Protected Resource Metadata (RFC 9728), a document that tells clients which authorization server it uses, and clients must use it to find that server.S4 §Overview The client then reads the authorization server's own metadata, published under RFC 8414 or OpenID Connect Discovery; clients must support both.S4 §Overview

  3. The client registers and the user signs in

    Before the flow starts, the client must get a client ID.S4 Clients and authorization servers should support Client ID Metadata Documents, in which a client names itself with a web address that points to its own description.S4 §Overview,2 Pre-registration is another route. Dynamic Client Registration is deprecated and kept only for authorization servers that do not support metadata documents.S4 The authorization server signs the user in if needed and issues the token.S4 §Roles The client must name the MCP server in a resource parameter on both the authorization and token requests.S4 When the response carries an issuer (iss) value, the client must check it against the issuer it recorded, before it sends the authorization code to a token endpoint.S4

  4. The client lists the tools

    From here every HTTP request carries the token in an Authorization: Bearer header, never in the URL's query string.S4 The server must check that the token was issued for it as the audience, and must not accept or pass on any other token.S4 The client sends tools/list, and the server returns each tool's name, description and input schema.S3 The list may vary with the authorization on the request, for example showing only the tools the caller's scopes permit.S3

  5. The model picks a tool and the client calls it

    Tools are model-controlled: the model can find and call them on its own, based on the context and the user's prompt.S3 The protocol does not require any particular way of involving the user, but there should always be a person who can deny a call, and applications should show confirmation prompts.S3 The specification's security principles also ask hosts to get the user's consent before invoking a tool, though MCP cannot enforce them.S1 The client then sends tools/call with the tool's name and arguments. The server returns the result and sets isError when the tool itself failed.S3

  6. The server asks for more scope when it needs it

    If the token lacks a scope an operation needs, the server should answer HTTP 403 with the error insufficient_scope and list every scope the operation needs in one challenge.S4 A client acting for a user should then re-authorize for the scopes it already asked for plus the new ones, so it keeps what it had.S4 Clients should limit their retries and stop after a few failed attempts.S4

03

Why does Model Context Protocol matter now?

Many MCP servers are reachable with no sign-in at all. In one academic scan, 3,233 of the 7,973 remote servers that answered took no authentication.3 Of 119 servers that let outsiders register an OAuth client, 114 accepted any redirect address, which lets an attacker collect an authorization code.4 At Black Hat, a research team showed that a malicious server picks the sign-in address a client opens, and turned that into code running as the user.5

The tool descriptions a model reads are a second way in. In the MCPTox study, poisoned descriptions for 353 tools on 45 real servers led agents to misuse tools they already held, at rates up to 72.8% for o1-mini.6

In September, five vendors shipped controls for MCP servers, and three of them route an agent's tool calls through an identity check.7 In a February concept paper, NIST's National Cybersecurity Center of Excellence listed MCP among the standards it is considering for agent identity.8

04

Where does it fit?

Compared withHow Model Context Protocol differs
OAuthOAuth is the framework underneath. MCP's authorization uses a chosen subset of OAuth 2.1 and related standards: a protected MCP server is an OAuth 2.1 resource server, and the MCP client is an OAuth 2.1 client.S4 §Roles
Dynamic Client RegistrationDynamic Client Registration lets a client register itself with an authorization server at run time. MCP's current specification deprecates it in favour of Client ID Metadata Documents and keeps it only for authorization servers that lack them.S4 §Overview
Cross App AccessCross App Access puts the company's identity provider between an agent and each application.9 MCP carries it as an opt-in extension, so an employee signs in once at the identity provider instead of approving each server.10
OpenPort ProtocolOpenPort is a separate proposal for exposing tools to agents through a server-side gateway, and it requires that an agent never see tools its credential cannot use.11 MCP only allows a server to vary its tool list by the caller's authorization.S3
05

What goes wrong?

  • Clients must treat tool annotations, the hints about how a tool behaves, as untrusted unless they come from a trusted server.S3 §Tool Tool descriptions can carry attacks too: poisoned descriptions have steered agents into misusing tools they already held.6
  • Authorization is optional in the specification,S4 which also says MCP cannot enforce its consent and access principles itself and asks implementers to build them.S1 In one scan, four in ten remote servers took no authentication.3
  • A server must not accept or pass on a token issued for anything else,S4 so a gateway that hands a person's token to the tools behind it is out of specification.12
  • Servers that accept anonymous registration with any redirect address let attackers collect authorization codes.4 The current specification has deprecated Dynamic Client Registration, the route those servers use.S4 §Overview
  • Servers must validate all tool inputs, implement proper access controls, rate-limit tool calls and sanitize tool outputs.S3 Clients should show a tool's inputs to the user before calling the server, to avoid malicious or accidental data leaks, and should log tool use for audit.S3
  • The specification makes the client a plain OAuth client. Nothing in the token says which person an agent acts for, and delegation is left to optional extensions.13
06

What should a security leader ask?

  1. 01

    Which MCP servers can our AI tools reach today, and which of them accept requests with no authentication?

  2. 02

    Does each MCP server we run check that a token was issued for it, and does any gateway in front of it forward a user's token to other services?

  3. 03

    Which of our MCP clients still register through Dynamic Client Registration, and when do they move to Client ID Metadata Documents?

  4. 04

    Which tool calls run without a person approving them, and where are tool inputs and calls logged for audit?

  5. 05

    When an agent calls a tool for an employee, can our logs show which person it acted for?

07

Sources

›Evidence: 13 claims from 10 briefings
  1. 1BI-FIND-2026-0069 · The rules cover one agent acting for one person; they break when that agent hands work to other agentsMCP with its community-added OAuth layer is the de facto tool-call protocol, and the panel is unanimous that it is not a policy system: it says how a client gets a token, not who is allowed to do what. The access decision is being pulled up into identity providers and agent gateways that sit above the protocol, which is exactly where Okta, Microsoft Entra and Ping Identity already sit for humans.
  2. 2BI-FIND-2026-0704 · Three vendors now route MCP tool calls through an identity check, and only Cloudflare's is generally availableOkta and Ping both shipped client ID metadata documents (CIMD), and CyberArk went the other way. With CIMD, an agent or app names itself with a URL that points to its own metadata, and needs no prior registration. Okta made CIMD a beta for AI agents and GA in preview for OIDC apps on September 16. Ping's Advanced Identity Cloud shipped it on September 29. CyberArk's September MCP feature relies on dynamic client registration (DCR) instead. Okta and Ping also changed their DCR support in the same releases, so neither has dropped it.
  3. 3BI-FIND-2026-0945 · 114 of 119 MCP servers that let outsiders register an OAuth client let them choose where sign-in codes are sentFour in ten of the remote MCP servers the authors found answered with no authentication: 3,233 of 7,973 servers that responded to an MCP handshake, against 2,428 using OAuth and 2,312 using static tokens or API keys. When the authors sampled the open ones, most were test or demonstration deployments with nothing sensitive in them. One was an internal customer-records server, meant to be private, that let anyone query more than 5,000 customer records (CVE-2025-61510). The count covers hosts that FOFA and Shodan index, so servers those engines miss are not in it.
  4. 4BI-FIND-2026-0944 · 114 of 119 MCP servers that let outsiders register an OAuth client let them choose where sign-in codes are sentOf the 119 remote MCP servers whose OAuth setups Zhou, Zhang and colleagues could test end to end, all of which register clients dynamically, 114 accepted any redirect address from an anonymous registrant. That lets an attacker register a client with a callback of their own, get the victim to click a genuine sign-in link, and collect the authorization code; seven of the nine CVEs the authors obtained trace to this flaw. Every server tested had at least one flaw, 325 in all. The 119 are what was left of 1,118 servers advertising a registration endpoint after the authors dropped duplicates, unreachable hosts and 316 servers behind enterprise access controls, so the rate describes servers open to outside registration, not MCP as a whole.
  5. 5BI-FIND-2026-0746 · An MCP server picks the login URL its client opens, and that was enough to run code on users' machinesMCP's sign-in flow lets the server tell the client where to send the user, and Zhong's team showed that the client passes that whole authorization URL to whatever opens it. A malicious or compromised server turned that into code running as the user: a PowerShell sub-expression through the widely used npm `open` package on Windows, which affected Google's Gemini CLI before version 0.17; file and SMB links that ran local or remote scripts; script injection in browser-based clients; and, in Anthropic's MCP Inspector, a stolen CSRF token that let a request start a local command. The fix in the official TypeScript SDK blocks javascript:, data: and vbscript: links but not file: or SMB, and the specification gives no rule for validating these URLs.
  6. 6BI-FIND-2025-0039 · MCPTox: a poisoned tool description steers most agents into misusing the tools they already holdA tool's description is enough to turn an agent against its user. MCPTox wrote poisoned descriptions for 353 tools on 45 real MCP servers and found agents acting on them across the board: o1-mini at 72.8%, DeepSeek-R1 at 70.9% and Phi-4 at 70.2%, down to 8.3% for Mistral, with refusals at 3% or below for every model. The attack needs no extra permission, because the poisoned text tells the agent to use a tool it already holds, such as reading an SSH private key or redirecting email to an attacker's address. The headline rates average three prompt settings, the paper's case counts do not add up, and its claim that more capable models are more vulnerable rests on one model family and breaks within it. No defense was tested.
  7. 7BI-FIND-2026-0703 · Three vendors now route MCP tool calls through an identity check, and only Cloudflare's is generally availableFive vendors shipped controls for MCP servers in September. Three of them proxy an agent's tool calls through an identity check, and only Cloudflare's proxy is generally available. Cloudflare's MCP server portals went GA on September 24. Okta's Agent Gateway APIs went to beta on September 16, and Google put MCP behind API Gateway in preview. The other two do different jobs: CyberArk registers MCP servers automatically when they support dynamic registration, and SailPoint lists them through its JDBC and REST connectors. The vendors agree that MCP servers need an identity control point. They do not yet agree on what it should do.
  8. 8BI-FIND-2026-0686 · NIST's agent identity work starts from OAuth, SPIFFE and SCIM and defers agents from outside the enterpriseNIST's National Cybersecurity Center of Excellence framed agent identity in February as a question of applying what exists, not inventing something new. Its concept paper lists OAuth, OpenID Connect, SPIFFE, SCIM, the NGAC access-control standard and MCP as candidates it is "currently considering", and names none of the agent-specific IETF work, such as ID-JAG, the WIMSE drafts or token exchange for agents. It keeps human delegation and logging in scope and leaves agents from outside the enterprise for a later effort. The next steps are a possible project description and a lab-built practice guide, with no date; the comment period closed on 2 April.
  9. 9BI-FIND-2026-0261 · Okta's answer to agent sprawl: put the identity provider between every agent and every appOkta's product chief presents Cross App Access as the company's answer for agents: the identity provider sits between an agent and each application, an administrator approves the connection once, and the agent receives a short-lived token with no consent prompt. An agent acting for a person gets the overlap of that person's permissions and its own granted scopes; an autonomous one gets its own identity. The mechanism is the identity assertion grant now in the IETF OAuth working group, and the Model Context Protocol carries it as an extension. Kahn claims more than twenty-five software vendors are adopting it and names only Anthropic.
  10. 10BI-FIND-2026-0678 · MCP's enterprise login extension says IdP revocation is immediate. Its own spec does not back that upMCP now has a stable way for a company's identity provider to stand between an employee's AI client and every MCP server. The client trades the employee's single sign-on ID token at the IdP for an ID-JAG, using OAuth token exchange, and hands that to the MCP server's own authorization server as a JWT assertion; the IdP checks group, role and conditional-access policy before issuing, and the employee sees one corporate login instead of a consent screen per server. The ID-JAG must be addressed to that authorization server, and when the client names a resource the access token must be restricted to it. The extension is opt-in: a server signals support in its authorization-server metadata.
  11. 11BI-FIND-2026-1239 · OpenPort holds agent writes as drafts for an operator by default; its own tests check less than its spec asksIn OpenPort, tool discovery is part of authorization. The manifest an agent fetches must not show tools or fields its credential is not allowed to use, and error details should be redacted so a denial does not reveal privileged names. The authors argue that tool-exposure protocols such as MCP often leave which tools are visible under which credentials to each server. Delegated OAuth tokens must be mapped to minimal OpenPort scopes and policy windows before any tool access, and the protocol names no identity provider: it "mandates that authorization decisions are made server-side and are auditable."
  12. 12BI-FIND-2026-0191 · MCP's authorization spec makes the agent a plain OAuth client and leaves delegation to an extensionThe specification binds every token to one server and forbids passing it on. Clients MUST send a resource parameter naming the server's canonical URI on every authorization and token request, and MUST check the issuer on every authorization response, the two controls against mix-up attacks. Servers MUST validate the audience of every token and 'MUST NOT accept or transit any other tokens'. Tokens travel only in the Authorization header, refresh tokens stay confidential and are never assumed, and offline access should not be advertised. That is the same anti-pattern the six-vendor agent draft names, a tool forwarding the token it received, and it is the rule a token-passthrough gateway breaks: a proxy that hands a person's token to the tools behind it is out of specification, whatever the vendor calls it.
  13. 13BI-FIND-2026-0189 · MCP's authorization spec makes the agent a plain OAuth client and leaves delegation to an extensionAs of the 28 July 2026 revision, an MCP client is an OAuth 2.1 client and an MCP server is an OAuth 2.1 resource server, and that is the whole identity model. Authorization is optional; HTTP transports SHOULD use it, and 'implementations using an STDIO transport SHOULD NOT follow this specification, and instead retrieve credentials from the environment'. The authorization server is out of scope. There is no actor claim, no on-behalf-of flow, no token exchange and no transaction-token profile: nothing in the token says whom the agent is acting for, which is what the Identiverse summit panel meant by 'not a policy system' and what WorkOS meant by the gateway being undefined. Anything of that kind is routed to the new extension model, optional and additive, in a separate repository. This page is the baseline against which the summit's forecast, that the specification gains a delegated-identity mechanism within a year, will be graded.

›In our briefings

The claim each of the newest briefings on Model Context Protocol would have you leave with.

  1. An MCP Events subscription inherits none of the limits on the access token that created it, which is short-lived, narrow in scope and checked on every request. The draft's design sketch keys a webhook subscription on the principal, the delivery URL, the event name and its arguments; the token, its scopes and its expiry are not part of that key. The principal MUST be authenticated and authorized at subscribe time, and after that the sketch says the server 'SHOULD periodically re-verify permissions'. WorkOS's Maria Paktiti: "There is no interval, no MUST, and no conformance test behind it." The lifetime is negotiable up to no expiry: ttlMs: null requests a subscription that never expires, and a server that grants it returns refreshBefore: null, so no refresh comes. Hence her line that 'the TTL you grant is your revocation window'. In her worked example, IT revokes a departing employee's access to a customer-contracts folder on Monday morning and the 24-hour subscription's next refresh lands Tuesday morning; a server that checked only at subscribe time, which is all the draft strictly requires, delivers every contract edit in between [figure 2].

    ChatGPT's MCP event subscriptions outlive the token that created them, and the draft sets no re-check scheduleBI-FIND-2026-0401

  2. The WorkOS CLI writes its MCP server into a coding agent only when the user opts in: the setup offer it may show after workos auth login or workos install defaults to No and stops asking once declined, a non-interactive, CI or --json run installs nothing, and opting in takes an explicit command such as workos mcp install or workos setup --yes. Once connected, the agent acts as the person who signed it in. Paktiti writes that the agent "inherits your dashboard role and can only do what your account is allowed to do." It works against one environment at a time, starts in a sandbox environment, and operates on production only when the user tells it to. Teams that want tighter limits get three independent admin settings (Enable, Allow production access and Allow write access), and all three are on by default. Apart from those settings and the sandbox default, the post describes no way to give the agent less than the user's own role.

    Opt in to WorkOS's MCP server and your coding agent acts as you, with write and production access on by defaultBI-FIND-2026-2401

  3. Uber runs per-hop token exchange with a full actor chain for AI agents in production, and has published sample tokens and a latency chart. Its security token service gives each hop a fresh token scoped to one audience, with a lifetime "in the order of minutes" (both sample payloads expire 300 seconds after issue), and the token carries the attested chain from the engineer to the agent making the call. Uber says "the system has been adopted by thousands of internal agents" and reports the exchange API's P99 as "consistently below 40 milliseconds" under current load; its chart sits mostly under 10 ms, with one spike reaching 40. The post covers this identity layer. Uber frames dynamic access control and a unified policy enforcement plane as its direction on top of it, and names session and intent claims as additions the token could carry in the future.

    Uber runs per-hop token exchange with a full actor chain for AI agents in productionBI-FIND-2026-0081

›Cited in 252 briefings

Every briefing that discusses this term, newest first.

  1. ChatGPT's MCP event subscriptions outlive the token that created them, and the draft sets no re-check schedule

    Article · · WorkOS Blog · Maria Paktiti · Oct 1, 2026

  2. Opt in to WorkOS's MCP server and your coding agent acts as you, with write and production access on by default

    Article · · WorkOS · Maria Paktiti · Sep 30, 2026

  3. Uber runs per-hop token exchange with a full actor chain for AI agents in production

    Article · · Uber Engineering blog · Matt Mathew et al. · May 21, 2026

  4. When the model cannot be trusted, control what the agent can reach, checked on every action

    Podcast · · Tech Talks Daily · AI Agent Security: Why Identity and Access Control Matter More Than Guardrails · Aug 11, 2026

  5. Auth0 ships an agent gateway in beta and moves the identity form into the chat so the agent never sees the card

    Video · · Okta · Uploaded September 24, 2026

  6. When AI agents hand work to other agents, the person who started the job can get lost

    Podcast · · The Non-Human & AI Identity Podcast · NHI & AI Identity Summit : Governing AI · Sep 7, 2026

  7. MCP's July spec ties client credentials to the authorization server that issued them

    Article · · Model Context Protocol project · Jul 28, 2026

  8. Singapore's regulator: each AI agent should have its own identity and, as a rule, no more access than its user

    Article · · Infocomm Media Development Authority (Singapore) · May 20, 2026

  9. AAuth gives every agent its own key, then keeps the delegation chain out of the token on purpose

    Draft · · IETF Datatracker · Dick Hardt, Hellō · Sep 25, 2026

  10. The JWT an agent or other OAuth client signs in with may name only the server's issuer, never the token endpoint

    Draft · · IETF Datatracker · Michael B. Jones et al. · Apr 28, 2026

  11. GitHub's AI triage agent got more consistent once code, not the model, ran the easy checks

    Article · · The GitHub Blog · Man Yue Mo et al. · Jan 20, 2026

  12. The fraud question may shift to: is this your agent, and what is it authorized to do?

    Podcast · · AI Security and the Law · Blurred Lines: How AI Is Rewriting the Rules of Fraud, Identity, and Defense · Aug 21, 2026

  13. Hassan's agent delegation draft now classes granting access as execute, not write

    Draft · · IETF Datatracker · Amr Hassan · Oct 22, 2026

  14. In an a16z survey of 100 CISOs, agent security climbs the 2027 list and access management slips

    Article · · Andreessen Horowitz · Stephenie Zhang et al. · Oct 28, 2026

  15. Palo Alto's own survey puts machine identities at 109 per person, up from 82 a year earlier

    Podcast · · Packet Protector · PP115: Palo Alto Networks: Reality of 109 to 1: Securing Machine Identities and AI Agents (Sponsored) · Jun 23, 2026

  16. OAuth token exchange does not guarantee a sub-agent stays inside its parent's access

    Article · · Control Plane by Karl McGuinness · Karl McGuinness · Mar 18, 2026

  17. OWASP's AIVSS lifts a CVSS 2.1 agent flaw to 7.1, and all ten worked examples land High or Critical

    Article · · OWASP AIVSS project · Ken Huang et al. · Oct 22, 2026

  18. MCP's November 2025 security guide makes proxy servers ask consent for each client before forwarding a login

    Article · · Model Context Protocol · Model Context Protocol project · Nov 25, 2025

  19. Backstage gives an agent a token that names only its user, so the agent can run anything the user can

    Video · · CNCF [Cloud Native Computing Foundation] · Uploaded November 10, 2026

  20. Sello has the services an AI agent calls sign its receipts; the agent can't fake them, but it can go around them

    Paper · · arXiv · Juan Figuera · Jun 2, 2026

  21. Woven by Toyota's exception agent recommends; a person decides, because a human has to answer for it

    Video · · Agentic AI Foundation · Uploaded November 4, 2026

  22. RBC's head of AI says most of its 50,000-plus agents are approved straight through

    Podcast · · Tech Disruptors · RBC on Turning AI Into a Competitive Advantage · Oct 20, 2026

  23. ceLLMate checks a browser agent's HTTP requests, not its clicks, and blocked all 12 scripted GitLab attacks

    Paper · · arXiv · Luoxi Meng et al. · Dec 14, 2025

  24. Okta and Google tie an agent's calls to the user, while GitHub's shared Copilot automations now act as Copilot

    Release tracker · · 11 vendors' release notes · Oct 2026

  25. MCP's client credentials extension puts the secret in the form body, while its server rule names HTTP Basic

    Article · · Model Context Protocol project · Oct 23, 2026

  26. A Nokia and China Mobile draft finds A2A and MCP stop at one hop, and wants cancellation carried down the chain

    Draft · · IETF Datatracker · Tirumaleswar Reddy.K et al. · Oct 27, 2026

  27. WIMSE's agent draft adds how an agent's access ends and where policy engines fit, without a new requirement

    Draft · · IETF WIMSE Working Group · Pieter Kasselman et al. · Sep 15, 2026

  28. OAuth's URL client ID draft tightens server rules in -03; the client is still whoever controls the URL

    Draft · · IETF Datatracker · Aaron Parecki et al. · Jul 6, 2026

  29. A digital ID is worth little until businesses can accept it, five voices say, and four point to bank onboarding

    Field signals · · 14 voices from the field · Jun 2024 to Oct 2026

  30. WorkOS's agent checklist limits a delegated agent to what it and its user may both do, checked on every call

    Article · · WorkOS Blog · Maria Paktiti · Sep 30, 2026

  31. JPMorgan won't let some agent sessions read and send email at once, says its security architect

    Podcast · · Insecure Agents · "Is This Tool Call Allowed? It's Never That Simple": Michael Davis (J.P. Morgan) · Sep 9, 2026

  32. Cisco's SD-Card rewrite pins MCP tools to what a registry signed; its security sections still read TODO

    Draft · · IETF Datatracker · Suhas Nandakumar et al. · Nov 1, 2026

  33. Cross-App Access draft warns that one stolen sign-in token could reach every app the identity provider permits

    Draft · · IETF Datatracker · Aaron Parecki et al. · May 21, 2026

  34. Give an agent access for the task and take it back afterwards, four voices say, and four say people approving each request cannot keep up

    Field signals · · 8 voices from the field · Feb 2026 to Oct 2026

  35. RFC 9207's issuer check stops at the redirect, and n8n's token-exchange flaw sat just past it

    Article · · WorkOS · Maria Paktiti · Jul 20, 2026

  36. A draft has each company sign its own half when one company's agent acts at another's request, refusals included

    Draft · · IETF Datatracker · Steven Mih, Action State Group · Oct 29, 2026

  37. AGTP checks that an agent ID is signed; binding it to the connection is an add-on in the author's patent notice

    Draft · · IETF Datatracker · Chris Hood, Nomotic, Inc. · Jun 28, 2026

  38. Each agent and each tool call can pass its check while the system fails, an Oxford-led survey argues

    Paper · · arXiv · Christian Schroeder de Witt · Apr 29, 2026

  39. BRICK, a proposed agent manifest, pins down which tools an AI agent may call but not who the agent is

    Draft · · IETF Datatracker · Samuel Sharaf et al. · Oct 26, 2026

  40. Agent URI draft -04 ties an agent's name to a key document and bars OAuth servers from trusting the name alone

    Draft · · IETF Datatracker · Yaswanth Narvaneni et al. · Oct 20, 2026

  41. Agent delegation records not anchored in trusted time become deniable once timestamps can be forged

    Draft · · IETF Datatracker · V.K. Uppalapati · Sep 29, 2026

  42. An agent audit log now names whom the agent acted for, but at low trust levels the agent may write that itself

    Draft · · IETF Datatracker · Raza Sharif, CyberSecAI · Oct 26, 2026

  43. Niyikiza's agent tokens narrow to a tool's arguments offline; revision -02 breaks -01 implementations

    Draft · · IETF Datatracker · Niki Aimable Niyikiza · Jun 15, 2026

  44. NovaFabric seals AI agent runs as signed evidence, and its author says a signature proves a key, not a person

    Paper · · arXiv · Mohsen Seyedkazemi Ardebili · Oct 25, 2026

  45. Approving a local MCP tool call does not limit the files it can reach, and researchers confirmed 150 cases

    Paper · · arXiv · Zeqian Huang et al. · Oct 15, 2026

  46. Receiver acknowledged, gateways cut off: in a fixed 16-host replay, 112 of 225 revoked tool calls still committed

    Paper · · arXiv · Lifei Liu et al. · Oct 11, 2026

  47. MCP's November 2025 authorization spec puts metadata documents ahead of dynamic client registration

    Article · · Model Context Protocol · Nov 25, 2025

  48. OAuth tokens prove an agent's rights, not a person's consent; revision 05 adds a day-long grant with spend caps

    Draft · · IETF Datatracker · Meiling Chen et al. · Oct 27, 2026

  49. A May 2025 agent-protocol framework, renamed twice and still unadopted, lists identity needs but no mechanism

    Draft · · IETF · Jonathan Rosenberg et al. · Jul 4, 2026

  50. OAuth scopes grant agents far more than the task needs, four voices say, and authorization handed from agent to agent has no settled pattern yet

    Field signals · · 7 voices from the field · Oct 2025 to Oct 2026

  51. A point-in-time access review can't keep pace with changing access, four voices say, and four vendors now put AI inside the review

    Field signals · · 9 voices from the field · Nov 2024 to Aug 2026

  52. Microsoft's Entra Suite demo blocks a planted prompt injection and an unsanctioned MCP server but gives no catch rate

    Video · · Microsoft Events · Uploaded November 24, 2025

  53. SailPoint’s JIT demo separates approved eligibility from expiring activation

    Video · · SailPoint · Oct 16, 2026

  54. MCP client setup moves from registration to metadata checks in a Spring walkthrough

    Video · · OWASP Foundation · Oct 24, 2026

  55. Shai-Hulud v4 used an MCP tool description to tell AI assistants to read SSH keys and keep quiet about it

    Video · · OWASP Foundation · Uploaded October 22, 2026

  56. IBM Research puts agent identity in a SPIFFE and OAuth sidecar; Maia Iyer calls authorization policy still open

    Podcast · · Kubernetes Bytes · Your AI Agent Has Too Much Access | Fixing Agent Identity · Sep 15, 2026

  57. Barclays mandates one gateway for its agents' budgets, blocks and audit, and its AI engineering director says the setup is not yet tested at scale

    Podcast · · Insecure Agents · The Fishbowl Problem: Agent Identity Across Clouds and Vendors, with Andy McMahon (Barclays) · Oct 5, 2026

  58. Agents and MCP servers get broad standing access because it is quick or easy, three voices say, and three vendors now pitch a gateway in front of MCP

    Field signals · · 9 voices from the field · Sep 2025 to Oct 2026

  59. AAuth carries the signals an agent's approval needs, Hardt says; the brain that decides does not exist yet

    Podcast · · Collateral Dialogue · Self-Driving Software: What OAuth Can’t Do for AI Agents | Dick Hardt, Hello · Sep 2, 2026

  60. The OWASP MCP Top 10 lead says few security teams can find the MCP servers developers run on their own machines

    Podcast · · The Application Security Podcast · Shadow MCP Servers: The AI Security Risk Hiding on Your Developers' Laptops · Oct 12, 2026

  61. At OWASP, one poisoned memory entry made a demo agent send a Q4 report to an attacker server

    Video · · OWASP Foundation · Recorded Jun 2026 · Uploaded October 21, 2026

  62. From Level 2 up, OWASP's AI security standard asks for an identity per agent and a check on every MCP tool call

    Video · · OWASP Foundation · Oct 21, 2026

  63. SaaS apps that only know users end up giving agents human accounts, says Okta's former chief architect

    Podcast · · The Identity Jedi Show: Where Identity Meets Strategy · Agents need a mission, not a key. · Oct 21, 2026

  64. Mythos also reports bugs that are not there, says Trend AI's CISO, so the work is in checking them

    Podcast · · Cloud Security Podcast · Mythos Hype vs. Reality: Why You Need an Agentic Harness · Oct 21, 2026

  65. Of 30 deployed AI agents, 7 publish how a website can recognize them and one signs its requests

    Paper · · arXiv · Leon Staufer et al. · Feb 19, 2026

  66. x402 saw about $1 million in agent payments in 30 days; Apify's founder says agents still buy through API keys

    Video · · AI Engineer · Uploaded October 18, 2026

  67. KuppingerCole's Mike Small blames most cloud incidents on customers' identity mistakes and sees CNAPP moving to cover AI

    Podcast · · Identity at the Center · #455 - CNAPP and the Cloud Identity Problem with Mike Small · Oct 19, 2026

  68. An IETF draft says agent access checks may refuse when a source is down, but must not record that as a 'no'

    Draft · · IETF Datatracker · Girish Konda · Oct 19, 2026

  69. Orange researchers would drop certificates from agent handshakes by putting post-quantum keys on a ledger

    Paper · · arXiv · Karol Budzyn et al. · Oct 16, 2026

  70. Six companies' agent-auth draft, since adopted by the WIMSE group, keeps credentials out of the model's reach

    Draft · · IETF Internet-Draft (individual submission, since adopted by the WIMSE working group) · Pieter Kasselman et al. · Jul 6, 2026

  71. Auth0 says its agent checkout token is limited to one purchase and amount; the tokens in its own diagram carry neither

    Article · · Auth0 Blog · Sep 28, 2026

  72. A quarter of 31,132 agent skills matched a risky pattern; 23 of the 87 riskiest showed clear signs of malice

    Paper · · arXiv · Yi Liu et al. · Jan 15, 2026

  73. Anthropic asks for per-agent identity; its guide leaves Claude Code sub-agent credentials unspecified

    Article · · Anthropic (Claude) · Anthropic · May 27, 2026

  74. Anthropic says agents should never use static API keys, yet its sub-agents can carry the parent's permissions

    Article · · Anthropic · Anthropic · May 18, 2026

  75. Zenity: one AgentCore agent's credentials reached all agents in its account and region; AWS calls it expected

    Article · · Zenity Labs · Tamir Ishay Sharbat · Oct 8, 2026

  76. Human Delegation Provenance is now an audit record only, and one that cannot prove it is complete

    Draft · · IETF Datatracker · Asiri Dalugoda, Helixar · Oct 6, 2026

  77. An end-to-end encrypted draft would hide an AI agent's task from the broker, but not who sent it or to whom

    Draft · · IETF Datatracker · Hamdy Atakora, Imara Labs · Sep 21, 2026

  78. Bu's checklist names a verifier for each agent-protocol claim and checks 'the same agent as before' separately

    Draft · · IETF Datatracker · Songbo Bu · Oct 14, 2026

  79. An attested OAuth client is one somebody vouched for; revision 12 says to trust its claims only as far as the voucher

    Draft · · IETF Datatracker · Tobias Looker et al. · Oct 13, 2026

  80. Okta's expired SCIM agent draft recorded certificates and OAuth clients; the joint draft after it dropped both

    Draft · · IETF Datatracker · Macy Abbey et al. · Oct 16, 2025

  81. With real MCP tools for cloud, database and email, most injected attacks got past four models, even with a plan

    Paper · · arXiv · Quan Zhang et al. · Mar 30, 2026

  82. A paper says trusting agents' self-reported quality reliably fails; its own test shows it in 10 of 36 settings

    Paper · · arXiv · Sunil Prakash · Mar 15, 2026

  83. OpenA2A's draft aims to keep secrets from the agent, and says no implementation yet supports its new controls

    Draft · · IETF Datatracker · Abdel Fane, OpenA2A · Oct 2, 2026

  84. Before an agent pays or changes data, Schrock's last check treats it as untrusted and keeps the record itself

    Draft · · IETF Datatracker · Iman Schrock, EMILIA Protocol · Sep 26, 2026

  85. Two implementations stop at the last of Reece's ten cross-company rules: a person approving one specific action

    Draft · · IETF Datatracker · Morgan Reece · Aug 31, 2026

  86. OpenA2A's protocol proves an agent holds a key and scores its behaviour, but never says who it works for

    Draft · · IETF Datatracker · Abdel Fane, OpenA2A · Oct 6, 2026

  87. Sharif's profile writes an AI agent's rights to pay and delete into an identity certificate valid up to a year

    Draft · · IETF Datatracker · Raza Sharif, CyberSecAI · Oct 2, 2026

  88. An MCP server picks the login URL its client opens, and that was enough to run code on users' machines

    Video · · Black Hat · Aug 20, 2026

  89. A coding agent tricked into writing one file lets the editor's language server send your keys out

    Video · · Black Hat · Aug 18, 2026

  90. Auth0's own Claude Code tutorial runs the agent on the developer's login and hands it fifteen admin scopes

    Article · · Auth0 Blog · Manish Hatwalne · Sep 29, 2026

  91. The official MCP TypeScript SDK now binds tokens to keys, ahead of the spec it implements

    Article · · WorkOS Blog · Maria Paktiti · Sep 29, 2026

  92. In Omada's survey of 577 identity leaders, executives report stronger AI agent controls than practitioners do

    Podcast · · The Security Strategist · Non-Human Identities and Agentic AI: The New Frontier in Identity Security · Apr 27, 2026

  93. Three vendors now route MCP tool calls through an identity check, and only Cloudflare's is generally available

    Release tracker · · 11 vendors' release notes · Sep 2026

  94. OWASP's agent Top 10 blames identity systems built for people, and does not say how it chose or ranked its risks

    Article · · OWASP GenAI Security Project · John Sotiropoulos et al. · Dec 9, 2025

  95. Okta says its agent kill switch revokes all tokens; its product screen says tokens outside the gateway survive

    Video · · Okta · Sep 24, 2026

  96. Okta's Oktane keynote: two agent products generally available now, support for other identity providers in 2027

    Video · · Okta · Sep 23, 2026

  97. Okta's September videos: six customer stories with few measured results, and keynote recaps that drop the ship dates

    Channel digest · · Okta (Oktane) · Sep 2026

  98. NIST's agent identity work starts from OAuth, SPIFFE and SCIM and defers agents from outside the enterprise

    Article · · NIST National Cybersecurity Center of Excellence · Harold Booth et al. · Feb 5, 2026

  99. MCP's enterprise login extension says IdP revocation is immediate. Its own spec does not back that up

    Article · · Model Context Protocol (modelcontextprotocol.io) · MCP project et al. · Jun 18, 2026

  100. A draft ties what each AI agent may read to what it may do, and still can't take a grant back

    Draft · · IETF Datatracker · Hamdy Atakora · Sep 21, 2026

  101. DIF's agent spec names every AI agent by a DID; the IETF's Agent Name Service switched to domain names

    Article · · Decentralized Identity Foundation (KYA-OS) · Dylan Hobbs et al. · Jul 29, 2026

  102. Microsoft limits stolen refresh tokens by restricting where they can be used, not by binding the token itself

    Podcast · · Entra.Chat · Stolen Tokens, Agents and Universal CAE: What Global Secure Access Adds to Entra ID · Oct 9, 2026

  103. Corvic's own tests find its agent vault keeps keys out of prompts but does much less against a misused connector

    Paper · · arXiv · Kenney, Corvic AI · Sep 27, 2026

  104. PayPal moved MCP tool permission checks into the server, and prompt-injected agents stopped getting through

    Paper · · arXiv · Huan Li et al. · Sep 18, 2026

  105. A survey rates only workload identity ready for agents, and argues no standard can tell what an agent intends

    Paper · · arXiv · Takumi Otsuka et al. · Apr 25, 2026

  106. A phishing agent hosted on Microsoft's own domain, and why giving agents identities does not stop it

    Article · · WorkOS Blog · Maria Paktiti · Sep 29, 2026

  107. Registering an MCP server on Bifrost ran its command, and the admin API that allowed it shipped with no password

    Article · · WorkOS Blog · Maria Paktiti · Sep 29, 2026

  108. AAuth's R3 lets an agent ask for an API's named operations, and a person approve one exact call

    Draft · · IETF Datatracker · Dick Hardt · Sep 28, 2026

  109. An integrator builds agent access on OpenID Connect and OAuth; clients want one view across many MCP gateways

    Podcast · · Identity Unfiltered · CIAM, Agentic AI & the Future of Identity · Oct 5, 2026

  110. Opal's access-review AI runs with a person approving by default, and an Opal engineer still reads the rejections

    Podcast · · Identity at the Center · #448 - Sponsor Spotlight - Opal Security · Sep 17, 2026

  111. Microsoft's Access Review Agent suggests whether to keep or remove access; the reviewer still decides, in Teams

    Podcast · · Entra.Chat · AI is Coming to Identity Governance! Meet the Entra Access Review Agent · Oct 18, 2025

  112. WorkOS's Tobin South, in November 2025: agents will get a new kind of account, not a new identity standard

    Podcast · · CodingCat.dev Podcast · 5.10 - How to Secure AI - Identity Access Management for Agents · Nov 21, 2025

  113. x402 v2 can take payment before delivering, with no identity beyond a wallet address and no way to dispute it

    Article · · x402 Foundation (GitHub) · x402 Foundation · Aug 31, 2026

  114. Cisco's SD-Card puts an A2A Agent Card in a selective-disclosure token, and leaves its registry out of scope

    Draft · · IETF Datatracker · Suhas Nandakumar et al. · Mar 2, 2026

  115. Eight panels agree an agent acts for a person, never as that person. When agents hand work to other agents, nobody had an answer.

    Conference · · Las Vegas · June 2026, a one-day summit track inside Identiverse 2026; the eight panels were published as podcast episodes between 18 August and 7 September 2026

  116. Three MCP auth flaws in a month, and the spec had already written the rule the SDKs skipped

    Article · · WorkOS Blog · Maria Paktiti · Oct 2, 2026

  117. OAuth handles an agent in one company, working while you watch. The OpenID paper says that is where it stops.

    Article · · OpenID Foundation · Tobin South et al. · Oct 2025

  118. Four vendors now say the access review is finished for agents. This one adds that agents will approve agents.

    Podcast · · The Non-Human & AI Identity Podcast · Ep # 15 - Identity is the 3rd Generation of Security · Sep 22, 2026

  119. The 'intent' in the Agentic JWT draft is a workflow step, and the agent reports its own fingerprint

    Draft · · IETF Datatracker · Abhishek Goswami · Jul 1, 2026

  120. The Agent Name Service draft was rebuilt on DNS, with a GoDaddy author, and no standards body has adopted it

    Draft · · IETF Datatracker · Scott Courtney et al. · Apr 13, 2026

  121. The IBAC paper's strict mode blocked every prompt injection, and its agent finished only a third of its tasks

    Article · · ibac.dev · Jordan Potti · Mar 3, 2026

  122. MCP made authorization optional, and two identity veterans say most servers took the option

    Podcast · · Hybrid Identity Protection Podcast · Agentic AI and the Authorization Gap No One Closed · Jun 9, 2026

  123. Two ways to give an AI agent an OAuth identity clashed in April; a draft now lets them work together

    Podcast · · Hybrid Identity Protection Podcast · Securing Non-Human Identities in the Age of Agentic AI with Sarah Cecchetti, Director of Product Management at Semperis · Apr 28, 2026

  124. Three vendors say AI systems now create identities on their own; the only survey behind it is one vendor's own

    Podcast · · The Data Exchange with Ben Lorica · The Rise of the Machine Identity: Securing the AI Workforce and AI Agents · Jan 29, 2026

  125. Okta's answer to agent sprawl: put the identity provider between every agent and every app

    Podcast · · Cloud Security Podcast · Who Governs Your AI Agents? Identity, Offboarding & Open Standards · Jul 2, 2026

  126. Auth0's agent gateway keeps every credential away from the agent; the post shows no step where a person approves

    Article · · Auth0 Blog · Andrés Aguiar · Sep 18, 2026

  127. An OpenID Foundation co-chair says AI agents need modest upgrades to today's identity systems, not a new one

    Podcast · · Identity at the Center · #390 - Identity Management for Agentic AI with Tobin South · Dec 8, 2025

  128. By Sierra's own account, its 75,000 agent sessions measure activity, not outcome

    Article · · Sierra · Neil Rahilly · Jul 9, 2026

  129. SecureAuth's sales chief sells agent identity on an unsourced ratio and on sums buyers work out from their own headcount

    Podcast · · GTM Vault · Securing AI Agents Beyond Authentication | Mark van Oppen, SecureAuth · Jul 5, 2026

  130. Mission-Bound OAuth's author now says the mission object does not belong in OAuth

    Article · · Control Plane by Karl McGuinness · Karl McGuinness · Mar 15, 2026

  131. Most of the work in Sierra's MCP gateway was identity: who each agent acts as and what it may reach

    Article · · Sierra Engineering · Mihai Parparita · Jul 22, 2026

  132. MCP's authorization spec makes the agent a plain OAuth client and leaves delegation to an extension

    Article · · Model Context Protocol project · Jul 28, 2026

  133. The first largely AI-run intrusion campaign got in by telling the model the attackers were a security firm

    Article · · Anthropic · Nov 13, 2025

  134. An MCP gateway can be demoed in a week; Sierra's rollout waited on the auth layer underneath it

    Article · · WorkOS Blog · Maria Paktiti · Sep 11, 2026

  135. Sierra's cloud agent runs 75,000 sessions a month and never holds a credential

    Article · · Sierra Engineering · Allen Chen et al. · Jul 15, 2026

  136. Highflame's open-source ZeroID narrows an agent's token at every hop, a step we did not find in Uber's write-up

    Article · · Highflame (GitHub) · Sep 25, 2026

  137. SailPoint's product chief wants an agent's access checked before each tool call and again at each resource access

    Podcast · · The Security Strategist · Who Owns an AI Agent? Rethinking Identity, Security, & Accountability · Aug 27, 2026

  138. Ping moved agent identity from roadmap to shipping product, and let AI agents run its admin tools too

    Conference · · Ping Identity · Austin · 2-day event (Day 1 / Day 2), 2026

  139. Canva sandboxes bounded conversion jobs; broad workforce agents pose a different access problem

    Podcast · · AI Security Podcast · Why I Dont Trust Your AI Agent | Kane Narraway, Canva · Aug 20, 2026

  140. xChk ships human OIDC portfolios and a separate agent approval API

    Paper · · arXiv · Sean MacGuire · Jul 15, 2026

  141. A July OAuth draft keeps agent identity stable as keys rotate

    Draft · · IETF Datatracker · Karl McGuinness · Jul 5, 2026

  142. In one author's MCP benchmark, a confirmation prompt on every call still let 6 of 10 attacks through

    Paper · · arXiv · Ting Liu, SymbolicLight · Jun 27, 2026

  143. OpenID's AuthZEN group approves a draft that lets policy tell an agent 'not yet, and here is what is required'

    Article · · OpenID Foundation · Serj Hallam · Jun 15, 2026

  144. In May 2026 Blackhawk's technology chief said Google's AP2 means little to merchants without a credential store

    Podcast · · So What About AI Agents · Agentic Commerce - Who's side is agent on? EP 60 - with Nik Sathe - Blackhawk Network (BHN) · Jun 10, 2026

  145. Agent payment protocols prove an agent may pay; a Toronto paper adds a step to check the job was done

    Paper · · arXiv · Adrian de Valois-Franklin et al. · Jun 7, 2026

  146. Huawei researchers add agent delegation to an existing permission model and prove each action traces to a human

    Paper · · arXiv · Amjad Ibrahim et al. · Jun 2, 2026

  147. Google's agent registry went GA in June while Okta, SailPoint and Ping pulled agents into their own lists

    Release tracker · · 11 vendors' release notes · Jun 2026

  148. Putting AIRGuard's policy in an agent's prompt barely cut attacks; enforcing it before each tool call did

    Paper · · arXiv · Suliu Qin et al. · May 27, 2026

  149. 114 of 119 MCP servers that let outsiders register an OAuth client let them choose where sign-in codes are sent

    Paper · · arXiv · Huijun Zhou et al. · May 21, 2026

  150. A proposed standard for agent authority advises against delegating it across company lines by default

    Paper · · arXiv · Partha Madhira · May 12, 2026

  151. Three vendors put MCP controls in three different places in May, and only Auth0's was GA

    Release tracker · · 10 vendors' release notes · May 2026

  152. An Imperial and Oxford protocol would give AI agents one approved use of a secret, never the secret itself

    Paper · · arXiv · Xiaohang Yu et al. · Apr 27, 2026

  153. A 'decentralized' agent identity draft blocks risky calls when its Registry is down, and did:aip is taken

    Draft · · IETF Datatracker · Paras Singla · Apr 17, 2026

  154. A proposed consent log for A2A and MCP records the terms an agent accepted, but the agent fills it in itself

    Paper · · arXiv · Ravi Kiran Kadaboina · Apr 16, 2026

  155. Most agent-payment protocols leave the agent's own reasoning out of scope, a survey of 142 works finds

    Paper · · arXiv · Qian'ang Mao et al. · Apr 15, 2026

  156. An agent certificate draft ties access to a trust score set by the issuing CA, usually the agent operator's own

    Draft · · IETF Datatracker · Raza Sharif, CyberSecAI · Apr 10, 2026

  157. AITH claims 4.7 million agent checks a second, but the test simulated its post-quantum signatures

    Paper · · arXiv · Zhaoliang Chen · Apr 9, 2026

  158. A proposed record of who authorized an AI agent's actions cannot yet prove which agent took them

    Paper · · arXiv · Asiri Dalugoda, Helixar · Apr 6, 2026

  159. An OAuth draft for AI agents' step-up login finishes the one-time-code path but leaves passkeys unfinished

    Draft · · IETF Datatracker · Martin Besozzi · Apr 3, 2026

  160. The commonest credential leak in AI agent skills is a debug print the framework hands to the model

    Paper · · arXiv · Zhihao Chen et al. · Apr 3, 2026

  161. Microsoft and Google made AI agent identity generally available in April; Okta's notes put its GA in May

    Release tracker · · 8 vendors' release notes · Apr 2026

  162. OpenID agent-claims draft adds a provider-set user, application or agent label; scores stay each provider's own

    Draft · · IETF Datatracker · Raza Sharif, CyberSecAI · Mar 26, 2026

  163. An agent-credential draft carries a hash of the user's instruction that its own verification steps never check

    Draft · · IETF Datatracker · Chudah Yakung, Attest · Mar 26, 2026

  164. AIP puts an agent's delegation chain in the token; its perfect score against attacks tests rules only AIP has

    Paper · · arXiv · Sunil Prakash · Mar 25, 2026

  165. A KubeCon demo moves MCP servers' OAuth checks into a gateway, safe only if nothing can reach a server directly

    Video · · CNCF [Cloud Native Computing Foundation] · Uploaded March 24, 2026

  166. Checking each agent call against the user's own task stopped all 634 off-task calls the authors wrote

    Paper · · arXiv · Reshabh K Sharma et al. · Mar 17, 2026

  167. An agent identity draft signs every tool call, but never finished saying whom the agent answers to

    Draft · · IETF Datatracker · James Cao et al. · Mar 16, 2026

  168. An MCP signing draft shows a client who wrote a tool, and admits a poisoned tool can still be validly signed

    Draft · · IETF Datatracker · Raza Sharif, CyberSecAI · Mar 14, 2026

  169. A survey labels 36.8% of agent skills malicious, but its own definition counts skills with any security flaw

    Paper · · arXiv · Yanna Jiang et al. · Feb 24, 2026

  170. A warning in the system prompt did not reliably stop coding agents obeying lines planted in a skill file

    Paper · · arXiv · David Schmotz et al. · Feb 23, 2026

  171. OpenPort holds agent writes as drafts for an operator by default; its own tests check less than its spec asks

    Paper · · arXiv · Genliang Zhu et al. · Feb 22, 2026

  172. Google DeepMind wants agent authority to narrow at every hand-off, and misreads MCP's scopes on the way

    Paper · · arXiv · Nenad Tomašev et al. · Feb 12, 2026

  173. MACAW Security's trust layer signs every agent call, but what stopped its prompt-injection test was a policy

    Paper · · arXiv · Mohan Rajagopalan et al. · Feb 11, 2026

  174. Auth0's AgentCore guide says prompts alone are not enough, then lets the prompt decide when to ask a person

    Article · · Auth0 Blog · Kapil Patil · Feb 10, 2026

  175. An open spec for stopping agent actions before they run needs the vendor's help to cover ChatGPT or Copilot

    Paper · · arXiv · Herman Errico · Feb 10, 2026

  176. A Cisco-led agent draft puts the verifiable credential at a public URL, with selective disclosure a SHOULD it does not yet show

    Draft · · IETF Datatracker · Jean Diaconu et al. · Feb 6, 2026

  177. 157 of 98,380 community agent skills proved malicious in a sandbox, and one publisher made more than half

    Paper · · arXiv · Yi Liu et al. · Feb 6, 2026

  178. A secure-MCP proposal registers every agent and tool centrally, and never mentions MCP's own OAuth rules

    Paper · · arXiv · Xinyi Hou et al. · Feb 1, 2026

  179. In February Okta let AI agents reach its management APIs over MCP; Cloudflare added controls on MCP connections

    Release tracker · · 7 vendors' release notes · Feb 2026

  180. Faramesh holds each agent action for a signed permit; its bypass tests assume no path around the gate

    Paper · · arXiv · Amjad Fatmi · Jan 25, 2026

  181. Rubrik ties agent recovery to protected backups and puts identity oversight in Agent Cloud

    Podcast · · AI Security Podcast · Why AI Agents Fail in Production: Governance, Trust & The "Undo" Button · Jan 23, 2026

  182. A proposed agent:// name survives a move between clouds but not a change of job

    Paper · · arXiv · Roland R. Rodriguez et al. · Jan 21, 2026

  183. Cisco Duo's MCP beta lets agents register themselves; Duo calls client ID metadata documents the future

    Article · · Cisco Duo · Colin Medfisch · Jan 7, 2026

  184. A Glasgow-Aarhus sandbox for MCP tools catches a secret printed as-is; base64 or truncation slips past it

    Paper · · arXiv · Zhuoran Tan et al. · Jan 3, 2026

  185. A Shenzhen paper ties each AI agent to a verified owner and its exact code, at 15 to 40 seconds of proving a step

    Paper · · arXiv · Zibin Lin et al. · Dec 19, 2025

  186. Oxford's VET checks an agent's output against its declared setup, and by the 32nd message four of five models cost over 3×

    Paper · · arXiv · Artem Grigor et al. · Dec 17, 2025

  187. A third-party MCP server can reconstruct what a user asked from the tool calls alone, in tests on simulated dialogues

    Paper · · arXiv · Yunhao Yao et al. · Dec 16, 2025

  188. Refuse the risky action when its plan relies on untrusted input, a Beihang survey of agent attacks proposes

    Paper · · arXiv · Guanquan Shi et al. · Dec 7, 2025

  189. No app, no developer: least privilege for an AI agent has to come from the task, six security researchers argue

    Paper · · arXiv · Mihai Christodorescu et al. · Dec 1, 2025

  190. One shared MCP token makes a whole company one identity in the logs; this paper wants per-user OAuth at a gateway

    Paper · · arXiv · Herman Errico et al. · Nov 25, 2025

  191. CORAL's public agent server rejected forged messages but let a wrong key open the session stream

    Paper · · arXiv · Yedidel Louck et al. · Nov 5, 2025

  192. Auth0's case for Token Vault: take the user ID out of the token call and let the session prove whose token it is

    Article · · Auth0 Blog · Abhishek Hingnikar, Okta · Nov 4, 2025

  193. On its authors' tasks, an LLM check of agent scopes works at one tool and refuses most valid three-tool requests

    Paper · · arXiv · Majed El Helou et al. · Oct 30, 2025

  194. AAGATE maps NIST's AI risk framework onto Kubernetes, and its paper never times the 'millisecond kill switch'

    Paper · · arXiv · Ken Huang et al. · Oct 29, 2025

  195. An MCP sandbox blocked every test attack that reached for the host; a poisoned tool that only redirected the agent got through

    Paper · · arXiv · Christoph Bühler et al. · Oct 24, 2025

  196. An agent OAuth proposal has the helper forward grants before receiving access tokens

    Draft · · IETF Datatracker · Kehan Yao, China Mobile · Oct 20, 2025

  197. Given 30 steps, the best computer-use agent solved four of 36 web hacking challenges, and newer Claude models scored lower

    Paper · · arXiv · Xiaoxue Ren et al. · Oct 14, 2025

  198. An October SCIM draft gives agents their own records, with security still unfinished

    Draft · · IETF Datatracker · Macy Abbey et al. · Oct 11, 2025

  199. A2AS promises prompt-injection defense at no latency cost, and publishes no test of it

    Paper · · arXiv · Eugene Neelou et al. · Oct 8, 2025

  200. An MCP threat survey shows agents trusting tool names and descriptions, and skips the spec's own OAuth rules

    Paper · · arXiv · Xinyi Hou et al. · Oct 7, 2025

  201. A decade of AI privacy papers: nearly half on training-data leakage, 2% on agents leaking user context

    Paper · · arXiv · Niloofar Mireshghallah et al. · Oct 2, 2025

  202. Toucan's 1.5 million MCP tool-use examples come only from servers that need no API key or account

    Paper · · arXiv · Zhangchen Xu et al. · Oct 1, 2025

  203. Left to run a decentralized-ID check, AI agents finished as few as 4 runs in 100; one pair agreed to skip half

    Paper · · arXiv · Sandro Rodriguez Garzon et al. · Oct 1, 2025

  204. Stytch sells a one-call check for whether a visitor is an AI agent, priced at half a cent a lookup

    Podcast · · So What About AI Agents · EP 39 - Agentic Authentication | Bobbie Chen · Sep 23, 2025

  205. Researchers found Claude Code's pull requests by a line the tool writes; maintainers merged 83.8% of them

    Paper · · arXiv · Miku Watanabe et al. · Sep 18, 2025

  206. Agentic JWT binds an agent's token to its task and its own code, but it is one author's draft, not a standard

    Paper · · arXiv (cs.CR preprint 2509.13597) · Abhishek Goswami · Sep 16, 2025

  207. A DeepMind paper puts agent markets on DIDs and credentials but leaves delegation unspecified

    Paper · · arXiv · Nenad Tomašev et al. · Sep 12, 2025

  208. A guide to securing agent plans has the planner choose each step's tools and leaves out agent identity

    Paper · · arXiv · Ron F. Del Rosario et al. · Sep 10, 2025

  209. A blockchain-agent survey's blueprint has no agent identity; risk scores and a human signature guard the wallet

    Paper · · arXiv · Nicolò Romandini et al. · Sep 8, 2025

  210. Tool poisoning runs through calls an agent is allowed to make; MindGuard looks for it in the model's attention

    Paper · · arXiv · Zhiqiang Wang et al. · Aug 28, 2025

  211. Twenty models nearly always name real MCP tools; the judge splits them on planning, which a schema check cannot catch

    Paper · · arXiv · Zhenting Wang · Aug 28, 2025

  212. MCPTox: a poisoned tool description steers most agents into misusing the tools they already hold

    Paper · · arXiv · Haohua Du · Aug 19, 2025

  213. A poisoned MCP tool got an agent to copy the client's config file without asking; deleting a file did ask

    Paper · · arXiv · Yongjian Guo et al. · Aug 18, 2025

  214. In a test harness, two of three MCP hosts refused a prompt asking for credentials; none refused a poisoned tool

    Paper · · arXiv · Yixuan Yang et al. · Aug 17, 2025

  215. A 'provable' bound on agent-memory injection assumes independent layers and fills in none of its detection rates

    Paper · · arXiv · Ken Huang et al. · Aug 17, 2025

  216. In x402 a wallet signature is the whole payment: no agent identity, no delegation, no refund once money moves

    Article · · x402.org (Coinbase Developer Platform) · Erik Reppel et al. · Aug 16, 2025

  217. MCP-Guard's classifier alone came within half a point of its best full pipelines' F1 on the authors' test set, and ran faster

    Paper · · arXiv · Wenpeng Xing et al. · Aug 14, 2025

  218. ERC-8004 stopped requiring agents to approve who rates them in January and admits fakes can inflate ratings

    Article · · Ethereum Improvement Proposals · Marco De Rossi et al. · Aug 13, 2025

  219. A national-lab provenance model logs an AI agent's prompts and decisions, but not whom it acted for

    Paper · · arXiv · Renan Souza et al. · Aug 4, 2025

  220. Sean O'Dell says Shared Signals audiences have stopped asking what CAEP is and started asking how to deploy it

    Podcast · · Identity at the Center · #361 - Identiverse 2025 - Sean O'Dell on Harnessing CAEP Abilities with Event-Driven Identity · Jul 14, 2025

  221. LPCI paper names instructions planted in agent memory that fire in a later session; its 43% failure rate matches a five-model average

    Paper · · arXiv · Hammad Atta et al. · Jul 14, 2025

  222. An MCP privilege study says obscure servers are riskier; its own table points the other way

    Paper · · arXiv · Zhihao Li et al. · Jul 5, 2025

  223. A July OAuth draft proposes authorization for models inside a client

    Draft · · IETF Datatracker · Yurong Song et al. · Jul 3, 2025

  224. A survey of MCP and A2A attacks pairs each identity threat with a fix, and tests none of them

    Paper · · arXiv · Mohamed Amine Ferrag et al. · Jun 29, 2025

  225. AI governance researchers argue OAuth tokens can't link an agent's work across services or sub-agents

    Paper · · arXiv · Alan Chan et al. · Jun 19, 2025

  226. MCP's June 2025 authorization spec: a server accepts only tokens issued for it and passes none upstream

    Article · · Model Context Protocol project · Jun 18, 2025

  227. RAS-Eval cites weak MCP authentication, yet maps none of its 3,802 agent attacks to an authentication weakness

    Paper · · arXiv · Yuchuan Fu et al. · Jun 18, 2025

  228. Each of eight AI models fell for at least one jailbreak planted in a simulated MCP server

    Paper · · arXiv · Junfeng Fang et al. · Jun 16, 2025

  229. A code scanner's top find in 583 MCP servers was hard-coded keys; tool poisoning needed an MCP-specific scanner

    Paper · · arXiv · Mohammed Mehedi Hasan et al. · Jun 16, 2025

  230. MCPWorld's 'MCP-only' agent also had a shell, and without it success falls from 53.23% to 43.00%

    Paper · · arXiv · Yunhe Yan et al. · Jun 9, 2025

  231. ETDI would sign MCP tool definitions against lookalikes, but leaves open who vouches for the publisher's key

    Paper · · arXiv · Manish Bhatt et al. · Jun 2, 2025

  232. In Microsoft's agent tests, label checks blocked policy-breaking tool calls and its FIDES planner finished more tasks

    Paper · · arXiv · Manuel Costa et al. · May 29, 2025

  233. AgentDNS gives AI agents one sign-in by having a central root hold each vendor's key

    Paper · · arXiv · Enfang Cui et al. · May 28, 2025

  234. A Cloud Security Alliance paper wants AI agents on a new identity stack, with OAuth kept for simple cases

    Paper · · arXiv (cs.CR preprint 2505.19301) / Cloud Security Alliance · Ken Huang et al. · May 25, 2025

  235. Function-calling models over-approve unsafe MCP calls; MCIP tests an offline classifier on synthetic chats

    Paper · · arXiv · Huihao Jing et al. · May 20, 2025

  236. Calling an MCP tool 'the best tool in the world' won the model's pick in almost every test

    Paper · · arXiv · Zihan Wang et al. · May 16, 2025

  237. The first Agent Name Service paper puts capabilities in the name and leaves protocol translation and governance out

    Paper · · arXiv (cs.CR preprint 2505.10609) / OWASP Agentic Security Initiative · Ken Huang et al. · May 15, 2025

  238. Decentralized AI agents have no off switch by default; builders say cutting API access still stops most of them

    Paper · · arXiv · Botao Amber Hu et al. · May 14, 2025

  239. Mastercard's Agent Pay began in April 2025 as a promise that payment agents would be registered and verified

    Article · · Mastercard · Mastercard · Apr 29, 2025

  240. An MCP gateway can do OAuth for servers that have none; this one passes them a user header

    Paper · · arXiv · Ivo Brett · Apr 28, 2025

  241. A 'zero trust' registry against AI tool squatting still has to trust its administrator

    Paper · · arXiv · Vineeth Sai Narajala et al. · Apr 28, 2025

  242. An A2A threat model flags the Agent Card as an attack path, and its example flow never passes on the user's identity

    Paper · · arXiv · Idan Habler et al. · Apr 23, 2025

  243. OWASP's multi-agent threat guide finds the hop where a user's permission check drops out, but names no fix

    Article · · OWASP GenAI Security Project · OWASP Agentic Security Initiative · Apr 22, 2025

  244. MCP Guardian wraps tool calls in token and regex checks, leaving poisoned servers and agent IDs open

    Paper · · arXiv · Sonu Kumar et al. · Apr 17, 2025

  245. Most of Progent's protection from prompt injection comes from a policy written before the agent reads anything

    Paper · · arXiv · Tianneng Shi et al. · Apr 16, 2025

  246. An early MCP security checklist was never tested, and the protocol has since adopted only part of it

    Paper · · arXiv · Vineeth Sai Narajala et al. · Apr 11, 2025

  247. In a 2025 MCP audit, Claude refused an encoded backdoor but wrote the plain one into .bashrc

    Paper · · arXiv · Brandon Radosevich et al. · Apr 2, 2025

  248. In Invariant Labs' Cursor test, hidden text in an MCP tool description got the agent to send out the user's MCP config

    Article · · Invariant Labs · Luca Beurer-Kellner et al. · Apr 1, 2025

  249. RFC 9728 tells clients where an API's tokens come from and leaves most of the trust decision to them

    Article · · RFC Editor · Michael B. Jones et al. · Apr 2025

  250. MAESTRO names your AI agents' identity threats and leaves designing the controls to you

    Article · · Cloud Security Alliance · Ken Huang · Feb 6, 2025

  251. An MIT-led paper has the user sign the agent's delegation token and lets OpenID Connect vouch for both parties

    Paper · · arXiv · Tobin South et al. · Jan 16, 2025

  252. RFC 8414 tells an OAuth client to discard server metadata whose issuer differs from the URL it started from

    Article · · RFC Editor · Michael B. Jones et al. · Jun 1, 2018

In our columns

  1. 114 of 119 MCP servers tested accepted any redirect. The registration debate misses it

    Column · Nov 13, 2026

  2. An agent's log must name whom it acted for, and the agent must not be the one writing it

    Column · Nov 9, 2026

  3. Switching off an AI agent needs a receipt from every system that held its access

    Column · Nov 7, 2026

  4. Signing MCP tools moves the risk to whoever registers them, so guard who may add a tool

    Column · Nov 2, 2026

  5. An AI agent's own trail can't show what it skipped; the gateway's can

    Column · Oct 16, 2026

Related terms

What appears alongside Model Context Protocol in the same briefings.

Cross App AccessOAuthOpenPort ProtocolAnthropicA2AClaude DesktopAgentic JWTTool squatting

Was this explainer useful?

Share this explainer

Pass it on. Every claim has a permanent number, so what you share today still points at the same words next year.