The original title is: "An AI agent exported a patient record. Your logs can't say who told it to."

Short summary

AI agents running under service accounts create an accountability gap: audit logs show the bot acted but not which human directed it, violating EU AI Act Article 12 requirements for identifying natural persons. The author built Crumb, a runtime gateway that stamps every agent action with an RFC 8693 delegated identity token linking the human subject to the agent actor, with multi-hop nesting for orchestrator chains and Sigstore-anchored tamper-evident logging. Crumb is a flight recorder for attribution, not a control plane, and its guarantees depend on the gateway being mandatory.

• •Service-account agent logs fail EU AI Act Article 12 by not capturing the human who directed each action
• •Crumb uses RFC 8693 token exchange to stamp delegated identity on every tool call, with nested act claims for multi-hop agent chains
• •Tamper-evidence is anchored to Sigstore's public Rekor log so operator rollback attacks are detectable

Generated with AI, which can make mistakes.