The problem. AI agents now execute tool calls that move money, change records, and touch production systems. Everyone can answer "who is the agent" (identity). Almost nobody can answer "what exactly did the agent do, and can a third party re-verify it?" Identity is necessary and not sufficient.
What we built. We wrapped an agent's tool-calling loop with a signing layer: every state-changing call gets an Ed25519 receipt bound to the call_id and the arguments digest. The spec is public (draft-correctover-ccs rev10, IETF), the verifier is self-serve (entropy.correctover.com), and the receipt schema is CC-BY-4.0.
The benchmark. We ran it against AgentDojo's banking scenario suite — 9 scenarios × 3 defense runs. Results, with the honesty up front:
-
27/27 runs: zero attack-goal success (
attack_goal_achieved=Truecount = 0) - 17/17 state-changing attack-path calls intercepted (DENY or ESCALATE), each with a verifiable Ed25519 receipt bound to the call
- Honesty clause you should demand from any benchmark post: in 10 of the 27 runs, the attack was aborted by the victim agent's own refusals before the interceptor was ever involved — those 10 runs are NOT credited to the interceptor. The 17/17 credit is exactly the runs where an attack path was launched and a state-changing call had to be stopped.
- 6 additional off-path DENYs (over-privileged calls outside the frozen task set), all receipted
- Full evidence: 605 files, 100 per-call receipts. Cryptographic freeze at GitHub tag
evidence-freeze-2026-11-19, timestamp witness via signed mail, public archive at Zenodo DOI 10.5281/zenodo.24356966 (CC-BY-4.0)
The boundary of the claim. This is a banking-suite-level result (n=1 suite, 1 victim, 1 attack family). We are not claiming general agent safety. What we do claim is the evidence shape: when something denies a call, is the receipt bound to call_id and args digest? Can a third party re-verify it offline? Those questions have a binary answer.
Why receipts, not logs. Runtime logs prove what the gateway saw. Signed receipts prove what anyone can re-verify without trusting the gateway. The difference matters the moment you have to answer to an auditor, a counterparty, or a court.
The public erratum culture, demonstrated on ourselves. Our first internal count of intercepted calls was 19. Source inspection showed 2 of them were off-path — the honest number is 17, and we publish the correction because a benchmark you can't reproduce is marketing, not engineering.
Reproduce it. One command against the frozen evidence pack: bash receipt-verify-demo.sh — two calls signed, sibling-call isolation PASS, independent verification PASS, pack stats printed (71 files / 100 receipts / 25 DENY-ESCALATE).
For teams building agent runtimes: if your platform issues action evidence of any kind (audit logs, receipts, attestations), we'd like to run a cross-verification exercise — your runtime, our verifier, or vice versa. Mutual verification beats mutual marketing.
Guigui Wang — Correctover (AI Reliability). Spec: draft-correctover-ccs (IETF). Verifier: entropy.correctover.com. Evidence: Zenodo 10.5281/zenodo.24356966.
Top comments (0)