A self-attested receipt is signed by the platform that ran the agent. An independently verifiable receipt is signed and anchored by an unaffiliated party, so any counterparty can check it without trusting the issuer.
The OWASP AI Exchange links draft-marques-asqav-compliance-receipts as an open profile for tamper-evident AI-agent logging under the EU AI Act, DORA, and US sector regulations.
AI coding agents open and merge their own pull requests. The git author string is mutable. A code-authorship receipt is a signed, chained, anchored record of who key-authored a change and when.
An agent that signs its own audit trail holds the signing key in process memory. Move the signer into a separate trust boundary and a compromised agent cannot forge receipts.
A tamper-evident log is one where an edit shows. The mechanics: signed receipts, a per-agent hash chain, external anchors, and verification that re-derives every hash.
One signed receipt can carry its own control mapping. How receipt fields provide verifiable evidence for NIST, OWASP, and NSA guidance without leaving the audit chain.
Snapshot the public key directory once, then verify ML-DSA-65 post-quantum receipts with no network access.
Shadow AI agents run unnoticed because there is nothing to scan. Where signing is enforced at an egress chokepoint, unsigned LLM traffic is either blocked or logged.
What our IETF Internet-Draft does, what each binding means, and why we anchor every receipt twice.
Public per-agent profile, embeddable badge, and an iframe you can drop into any repo. Buyers verify signatures themselves.
On-demand COSE_Sign1 export over ML-DSA-65. Submit to a SCITT transparency service and parse with cbor2.
npm install @asqav/sdk. Same ML-DSA-65 backend as the Python SDK. Wrap a Vercel AI SDK or LangChain.js tool call.
Logs are not evidence. For EU AI Act Article 12, DORA, SOC 2 you need tamper-evident, portable records.
pip install asqav[cli] then asqav demo. Four scenarios, signed receipts, zero setup.
Reference implementation of governance metadata for A2A Agent Cards: trust_score, retention TTL, derivation rights.
Signed governance attestations let any third party verify an agent's trust level offline. L0 to L3 canonical mapping.
Most governance tools sign after the fact. Three-phase captures intent, decision, and execution separately.
trace_id links every action in a multi-agent workflow. emergency_halt() revokes every agent instantly when something goes wrong.
Scope tokens prove what an agent can do. Nonce-based replay protection makes each token single-use.
Auth tokens prove identity. Scope tokens prove permission. Verifiable by any receiver without calling your org.
Reconstruct and verify agent action timelines from signed receipts. Online or offline.
Test governance rules safely with observe mode, semantic action patterns, and plain-English explanations.
Package your audit trail into a self-contained compliance bundle with Merkle root verification.
Six capabilities every organization needs before EU AI Act high-risk enforcement (Annex III deferred to December 2027 under the pending Digital Omnibus).
Three lines of code to add audit trails and policy enforcement to any LangChain agent.
One compromised key means one rogue agent. Multi-party signing mitigates single points of failure.
MCP gives agents unlimited tool access. Your security model needs to catch up.
NIST standardised ML-DSA in FIPS 204. Why long-retention audit trails should sign with it.
The security questions your board will ask next quarter. Get ahead of them now.
Articles 12, 19, and 26 decoded. What your audit trails actually need to contain.
Logs record what happened. Cryptographic proof makes it undeniable.
How DORA applies to AI agents in financial services and what ICT risk management requires.
Lattice-based signatures, key sizes, and performance. The practical guide to FIPS 204.
Get practical insights on AI agent security and compliance obligations. No spam, unsubscribe anytime.