@ceyptoslim2
Tavares HeathFounder▲6
Founder, Independent
4d ago · edited · AI Engineering, AI Adoption, Agents, Testing and Evals · Miami

I fixed the gap

Three days ago a cybersecurity professor critiqued my project in public: an agent-security audit log is self-attestation — the system investigating itself writes its own evidence. He was right. As of tonight it's closed.

Chain heads are now anchored to custody sinks the system can append to but never rewrite — a git repo plus RFC 3161 timestamps — with verification that re-walks the store against the anchors without trusting the system. A rewrite-and-rehash attack (modify a record, recompute every hash consistently so the internal verifier passes green) is caught by the external anchors as a contradiction.

That exact attack is executed by CI, and the run records are frozen in the repo. Critique, design, implementation, attack, detection — all public, all this week.

If governance for autonomous agents is your lane, I'm looking for design partners who'll try to break it.
github.com/factnotebook/public-reconstructions

AgentsSecurityGovernance

Comments2

  • @ryanallis
    Ryan Allis
    Founder of AI Lounge and SaasRise, helping founders reach $100M+ exits
    3d ago

    Fixing a professor's public critique in three days, all in the open, is a story security buyers will remember.

  • @ceyptoslim
    Tavares Heath
    Policy-governed execution layer for AI systems. Enforces deterministic decisions and produces verifiable audit evidence.
    2d ago

    Thanks Ryan — the three-day part is the least interesting bit. What makes it stick is that the fix wasn't self-reported: an independent researcher reconstructed the evaluation and published the verification (the link in the post is his registry, not my repo). A story security buyers can check is the only kind they remember.

    And the door's open for the second half — anyone who wants to try to break the gate rather than read about it, that's the design-partner conversation I'm after.

Log in to join the thread.