Skip to content
IAM Insights

Explainers · Glossary

Prompt injection

prompt-injected · prompt manipulation · indirect prompt injection

Hiding instructions in text an AI model reads, such as an email or a web page, so the model does something its operator never asked for.

In our briefings

The claim each of the newest briefings on Prompt injection would have you leave with.

  1. Neiderman's case that the identity stack has to be rebuilt for agents starts with history. Identity spent decades as IT plumbing behind the firewall, the cloud made the identity provider the real gatekeeper, and companies answered with point tools: he recalls an average of about 8.3 per company, from a spreadsheet of conversations with, by his estimate, 100 to 200 or more CISOs. Agents, he says, will bring roughly a hundred times more identities working around the clock, and probably 99.9 percent of actions on data and systems will be theirs. In his view a stack built for humans and carrying decades of undocumented connections will not hold that load, while NewCore, starting fresh, can adjust its design as agent protocols are "still being formed". Two less interested voices in other sources we have read make a similar case. The founder of an identity threat detection company Okta had just acquired warned that bolting a bunch of things on will fail over time, and a paper led by Cloud Security Alliance authors lays out a new layered identity architecture for agents.

    Can your identity stack be patched for AI agents, or does it have to be replaced? NewCore's CTO says replacedBI-FIND-2026-0103

  2. Linx's Goldenberg calls one agent acting for one user "in a way" a solved challenge, because the user's token usually already carries what downstream applications need to control data access. Once agents delegate to other agents across several hops, he says, "you lose track of who is the initial actor", and he sees toxic combinations and violations of segregation of duties. Aembit's Goldschlag answers with token exchange at every hop, from user to agent to gateway to MCP server to resource, each hop holding only the credential for the next step. Godha points to IETF drafts on chained delegation tokens and to transaction tokens, which he says are coming.

    When AI agents hand work to other agents, the person who started the job can get lostBI-FIND-2026-0070

  3. Asked whether a review of access every three months still applies to agents, Alexis Moyse, CEO of Clarity Security, says it doesn't really apply: a short-lived agent's task may last 15 seconds, and its lifecycle begins and ends with the request. Moyse's stated goal is "to eliminate access reviews altogether". Moyse says reviews of people stop mattering too once their access is just as short-lived, and describes the ideal as access granted for a task and taken away the moment it is done. Itamar Apelblat of Token Security says the point leaves him asking whether access reviews should exist at all in a world this dynamic. Homayun Yaqub of Digital Hands calls the shift dynamic governance, but describes it as "a conceptual model that we need to explore" and says "I'm not sure we're quite there yet", mostly because what defines an agent's lifecycle is still being learned.

    Clarity's CEO says quarterly access reviews don't really apply to 15-second AI agentsBI-FIND-2026-0086

›Cited in 117 briefings

Every briefing that discusses this term, newest first.

  1. Can your identity stack be patched for AI agents, or does it have to be replaced? NewCore's CTO says replaced

    Podcast · · TechSpective Podcast · Rebuilding Identity Security From Scratch for the Age of AI Agents · Sep 21, 2026

  2. 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

  3. Clarity's CEO says quarterly access reviews don't really apply to 15-second AI agents

    Podcast · · The Non-Human & AI Identity Podcast · NHI & AI Identity Summit : Where Traditional IAM Breaks · Aug 28, 2026

  4. Microsoft wants AI agents governed like employees, each with a human sponsor

    Article · · Microsoft Security Blog · Joy Chik · Jan 20, 2026

  5. 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

  6. Knowing who an agent is and what it may reach is not enough, six voices say: check each action as it runs

    Field signals · · 14 voices from the field · Dec 2025 to Nov 2026

  7. A user's approved request is not yet authority an agent can be held to, and most deployments skip that step

    Article · · Control Plane by Karl McGuinness · Karl McGuinness · Mar 17, 2026

  8. 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

  9. 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

  10. Wise wants customers to approve one locked job, so a tricked agent's payment fails at the service check

    Video · · AI Engineer · Uploaded November 5, 2026

  11. An OAuth draft lets the API, not the agent, describe the approval, and now name the key the token is bound to

    Draft · · IETF Datatracker · Yaroslav Rosomakho et al. · Jun 25, 2026

  12. NatWest found its LLM judge barely tracked human reviewers, and an 83% pass rate hid 43% on topic changes

    Video · · AI Engineer · Uploaded November 4, 2026

  13. Malware now tells AI triage tools to ignore it, and in Talos's test the bluntest request worked best

    Article · · Cisco Talos Blog · Ryan Fetterman · Oct 8, 2026

  14. Anthropic says Claude Code's auto mode lets 17% of overeager actions through, mostly by misjudging consent

    Article · · Anthropic · John Hughes · Mar 25, 2026

  15. 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

  16. Raut's revised agent draft keeps token exchange out of the model's reach, and concedes the token carries no evidence of it

    Draft · · IETF Datatracker · Ashay Raut · May 22, 2026

  17. Contractors mimicked Lovable's coding agent, its engineers say; one wants agents' changes logged as the agent's

    Podcast · · Cloud Security Podcast · Securing Millions of AI-built Apps: How Lovable Defends Against Insider Threats · Sep 11, 2026

  18. IETF workload practices draft describes single-audience JWTs in place of stored secrets and discourages token exchange

    Draft · · IETF Datatracker · Arndt Schwenkschuster et al. · Sep 22, 2026

  19. 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

  20. 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

  21. Attackers treat AI keys as production credentials, Anthropic's September misuse report says

    Article · · Anthropic · Anthropic Threat Intelligence team · Sep 10, 2026

  22. 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

  23. 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

  24. Textile maker's CIO and CISO has no way to find shadow AI, and calls an agent its own kind of identity

    Podcast · · Identity at the Center · #446 - Rethinking Identity for AI Agents with Rick Scot · Sep 7, 2026

  25. 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

  26. 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

  27. 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

  28. 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

  29. 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

  30. An undotted configuration key and a broad workload identity token let a GitHub issue reach a Gemini CLI cloud project

    Podcast · · CyberWire Daily · The wrong kind of access. [Research Saturday] · Oct 17, 2026

  31. 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

  32. 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

  33. 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

  34. 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

  35. AgentArmor cut attack success to 3% in its authors' tests; most misses came when a user said to follow a file

    Paper · · arXiv · Peiran Wang et al. · Nov 18, 2025

  36. An OAuth profile for AI agents makes every hand-off cut privileges, but a shorter expiry alone counts as a cut

    Draft · · IETF Datatracker · Angel Cruz · Feb 7, 2026

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

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

  38. 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

  39. 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

  40. 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

  41. 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

  42. OpenID experts tell EMVCo an agent's payment can look like the cardholder's, and no one vouches for the agent

    Article · · OpenID Foundation · Serj Hallam et al. · Oct 8, 2026

  43. 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

  44. A2A 1.0 connects agents across companies and leaves out the person they act for

    Article · · A2A Protocol (Linux Foundation) · A2A Protocol project et al. · May 28, 2026

  45. 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

  46. NIST's request on securing AI agents asks about hijacking and rollback, and never about who the agent is

    Article · · Federal Register (govinfo) · NIST Center for AI Standards et al. · Jan 8, 2026

  47. 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

  48. 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

  49. 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

  50. Two IETF veterans want every specification to carry guidance for AI agents, with authorization kept outside the model

    Draft · · IETF Datatracker · Orie Steele et al. · Sep 11, 2026

  51. A SANS instructor's broker vaults agents' real keys, but its third delivery model breaks the abstract's promise

    Draft · · IETF Datatracker · Kenneth G. Hartman · Mar 29, 2026

  52. 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

  53. 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

  54. 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

  55. 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

  56. 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

  57. The directors' own handbook now puts one question about AI agent identity to boards

    Article · · National Association of Corporate Directors and Internet Security Alliance · Larry Clinton et al. · Mar 12, 2026

  58. Only 18% of firms give AI agents their own access, and 68% cannot tell an agent's action from a person's

    Article · · Cloud Security Alliance blog · Cloud Security Alliance · Apr 20, 2026

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

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

  60. 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

  61. 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

  62. A preprint cuts off every sub-agent within about 40 seconds of its parent stopping, with no revocation server

    Paper · · arXiv · Saurabh Deochake, SentinelOne · May 20, 2026

  63. The same model took out-of-scope actions in 1.1% of runs in OpenHands and 27.7% in Claude Code

    Paper · · arXiv · Yubin Qu et al. · May 18, 2026

  64. Clicks and timing alone told Oxford researchers which of 14 models was driving a browser agent

    Paper · · arXiv · William Lugoloobi et al. · May 14, 2026

  65. A Singapore demo holds an agent's stablecoin payment until compliance clears, but says nothing on agent identity

    Paper · · arXiv · Kenneth See et al. · Apr 30, 2026

  66. 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

  67. 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

  68. 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

  69. 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

  70. 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

  71. 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

  72. A Huawei agent draft keeps one 'master agent' as gatekeeper and drops the line that had servers re-check it

    Draft · · IETF Datatracker · Yuan Ni et al. · Feb 28, 2026

  73. A draft binds workload credentials to their host via a local agent, and leaves that agent's compromise undecided

    Draft · · IETF Datatracker · Ram Krishnan et al. · Feb 26, 2026

  74. 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

  75. 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

  76. eBay researchers say AP2 does not make a payment mandate single-use, so they built a gateway that does

    Paper · · arXiv · Qianlong Lan et al. · Feb 6, 2026

  77. AgentGuardian's learned rules caught 18 of 20 misuse cases; its sample rule, as printed, allows any text file

    Paper · · arXiv · Nadya Abaev et al. · Jan 15, 2026

  78. 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

  79. A "don't ask again" for Python let a tampered Agent Skill upload slides from Claude Code

    Paper · · arXiv · David Schmotz et al. · Oct 30, 2025

  80. 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

  81. Attacks tuned to each defense got past all 12 prompt-injection and jailbreak defenses tested

    Paper · · arXiv · Milad Nasr et al. · Oct 10, 2025

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

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

  83. 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

  84. 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

  85. 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

  86. 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

  87. 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

  88. The Aegis agent-security paper's 2.79-second proof time is a benchmark-based target its simulation reports back

    Paper · · arXiv · Sai Teja Reddy Adapala et al. · Aug 22, 2025

  89. 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

  90. 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

  91. 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

  92. A Tsinghua blockchain design for AI agents backs its defense claims with a table and reports no attack run

    Paper · · arXiv · Zhenhua Zou et al. · Aug 2, 2025

  93. ChatGPT agent’s launch separated connector reads from website-action login

    Article · · OpenAI · OpenAI · Jul 17, 2025

  94. Meta's injection-trained Llama kept agent attacks to 2.1% or less, but its authors tried no attack tuned against it

    Paper · · arXiv · Sizhe Chen et al. · Jul 3, 2025

  95. 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

  96. 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

  97. 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

  98. Meta's LlamaFirewall cut agent attacks to 1.75% in a replay that counts late catches as defended

    Paper · · arXiv · Sahana Chennabasappa et al. · May 6, 2025

  99. 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

  100. 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

  101. Google's CaMeL stops nearly all prompt injection in tests, but agents finish fewer tasks and need written policy

    Paper · · arXiv · Debenedetti · Mar 24, 2025

  102. Splitting an agent's privileges stopped every tested prompt injection, if users refuse each alert

    Paper · · arXiv · Juhee Kim et al. · Mar 17, 2025

  103. A fake error message relayed between agents got three multi-agent frameworks to run an attacker's code

    Paper · · arXiv · Harold Triedman et al. · Mar 15, 2025

  104. All eight prompt-injection defenses in this test were bypassed by attacks tuned to each one

    Paper · · arXiv · Qiusi Zhan et al. · Feb 27, 2025

  105. Testing AI agents one at a time misses attacks that combine them, a Cooperative AI Foundation report argues

    Paper · · arXiv · Lewis Hammond · Feb 19, 2025

  106. A Carnegie Mellon defense blocked every policy-breaking injection in tests, if its labels and tools hold up

    Paper · · arXiv · Peter Yong Zhong et al. · Feb 13, 2025

  107. NIST's best new attack hijacked a Claude agent on 81% of test tasks, up from 11%; retries helped attackers too

    Article · · NIST Center for AI Standards and Innovation · CAISI technical staff · Jan 17, 2025

  108. 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

  109. SecAlign held optimised prompt injections to 14% or less on five models; its authors say it cannot stand alone

    Paper · · arXiv · Sizhe Chen et al. · Oct 7, 2024

  110. A 2024 USENIX benchmark found none of ten prompt-injection defenses sufficient, its best detector included

    Article · · USENIX Security 2024 · Yupei Liu et al. · Aug 14, 2024

  111. The 2024 'IDs for AI' paper puts the ID on a chat session, not a model, and leaves trust in its issuer to others

    Paper · · arXiv · Alan Chan et al. · Jun 17, 2024

  112. OpenAI's instruction hierarchy ranks who may give orders by where the text arrived, not by who sent it

    Paper · · arXiv · Eric Wallace et al. · Apr 19, 2024

  113. No spotlighting test spans the abstract's 50%-to-2% drop; the one that does is a few-shot run the authors doubt

    Paper · · arXiv · Keegan Hines et al. · Mar 20, 2024

  114. IsolateGPT walls off each app an AI assistant uses, and its attack results assume the user refuses every warning

    Paper · · arXiv · Yuhao Wu et al. · Mar 8, 2024

  115. StruQ held hand-written prompt injections to 2% or less on two small models; optimised attacks still got through

    Paper · · arXiv · Sizhe Chen et al. · Feb 9, 2024

  116. In 2023 OpenAI floated agent IDs that name an accountable user, and left open how anyone would verify one

    Article · · OpenAI · Yonadav Shavit et al. · Dec 14, 2023

  117. The 2022 'ignore previous prompt' paper hijacked GPT-3 58.6% of the time, with no tools or agent in the test

    Paper · · arXiv · Fábio Perez et al. · Nov 17, 2022

Related terms

What appears alongside Prompt injection in the same briefings.

Instruction hierarchySpotlightingWASPMicrosoft Entra Secure Web and AI GatewaySkill-based prompt injectionAgent delegation