Skip to content
IAM Insights

Explainers · Glossary

Side-channel attack

side channel · side-channel · side channels · timing side-channel

Learning a secret from indirect signals, such as timing or whether a step ran, instead of reading the secret itself.

In our briefings

The claim each of the newest briefings on Side-channel attack would have you leave with.

  1. Under this draft a client's identity is a web address and whatever its owner publishes there. The client_id is an https URL, which revision -03 says "MUST contain a path component other than /", and the JSON document at that URL "MUST contain a client_id property whose value MUST match" it. The authorization server learns the client's name, redirect addresses and keys by fetching that document. Its security section says a client that changes its URL is "a brand new client with no relationship to the previous one", and that losing control of the domain would let a third party "assume the client's identity". We read this as a name anyone can look up. That is more than an anonymous registration gives a server, though it says nothing about what software sits behind the name.

    OAuth's URL client ID draft tightens server rules in -03; the client is still whoever controls the URLBI-FIND-2026-0319

  2. Datadog's goal for its code sandboxes, in the speakers' words, is "to ensure that a sandbox may use credentials, but credentials never end up in a sandbox." In the design they describe, the agent harness runs inside the sandbox and reaches GitHub or the AI provider through a network gateway. When a sandbox is created, Datadog's internal identity service gives it a new identity that, the speakers say, "only allows the sandbox to authenticate to the gateway. Nothing else." The gateway checks that this identity may reach GitHub and the specific repository, fetches the token from a secret store, forwards the request and strips credentials from the response. It can also filter which API calls pass, for example letting a sandbox open a pull request but not approve one; the speakers note that GitHub gives both actions the same credential scope. They do not say what form the sandbox identity takes.

    Datadog keeps tokens out of its AI agents' sandboxes: each holds an identity only the gateway acceptsBI-FIND-2026-1568

  3. RFC 9901 never mentions agents or delegation. Its holder "may refer to the actual user, the supporting hardware and software in their possession, or both", and section 9.9 covers a credential moving past its holder: any entity holding an SD-JWT can forward it to any third party that does not enforce key binding, and may remove disclosures as it does. The RFC says this "could be desirable" in some cases and gives a supply chain where each party can choose to pass on less. The set of disclosures in a bare SD-JWT is not integrity protected, so the receiver has no way to tell what was dropped, and, as we read the format, nothing in it names who forwarded it. An agent handing on authority under the bare RFC therefore looks like whoever presents the token, which is why the delegation designs in our briefings add a chain of their own; our briefing on the Delegate SD-JWT draft found it lets the key-binding JWT itself be an SD-JWT that names the next holder's key.

    The SD-JWT standard that agent drafts build on never mentions agents and leaves key binding to each verifierBI-FIND-2025-0143

›Cited in 11 briefings

Every briefing that discusses this term, newest first.

  1. 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

  2. Datadog keeps tokens out of its AI agents' sandboxes: each holds an identity only the gateway accepts

    Video · · fwd:cloudsec · Uploaded September 25, 2026

  3. The SD-JWT standard that agent drafts build on never mentions agents and leaves key binding to each verifier

    Article · · IETF (RFC Editor) · Daniel Fett et al. · Nov 20, 2025

  4. With filters off, Mythos Preview and GPT-5.5 exploited 157 and 120 of 898 real bugs; defenses stopped most

    Paper · · arXiv · Zhun Wang et al. · May 11, 2026

  5. Code agents in the CIBER benchmark resisted deleting sensitive files far better than copying or reading them

    Paper · · arXiv · Lei Ba et al. · Feb 23, 2026

  6. Wisconsin researchers check each agent action against its history, in small tests with hand-written rules

    Paper · · arXiv · Nils Palumbo et al. · Feb 18, 2026

  7. 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

  8. 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

  9. 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

  10. 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

  11. The 2023 paper that named indirect prompt injection showed an attacker can use whatever access the assistant has

    Paper · · arXiv (cs.CR) · Kai Greshake et al. · Feb 23, 2023