AI Audit Agent Uncovers Seven Bugs in Cloudflare Crypto Library

AI audit agent zkao spotted seven vulnerabilities inside Cloudflare CIRCL, a widely used cryptography library.

The agent combined models from Anthropic and OpenAI with specialized verification skills. It flagged two critical issues on its own before human review.

One bug involved precision loss during threshold RSA operations. The second broke access control checks in an attribute-based encryption scheme.

Both problems sat in production code for months. They were confirmed through HackerOne reports and fixed before public disclosure.

This case shows how automated agents can now handle large parts of vulnerability discovery in complex cryptographic code.

Agent locates the first set of issues

zkSecurity ran zkao across the full CIRCL codebase for multiple weeks. The agent surfaced a total of seven distinct problems. Four of them required follow-up human confirmation while three were verified automatically.

The precision-loss bug appeared when floating-point arithmetic was used inside threshold RSA. Small rounding errors could corrupt signature shares in certain edge cases. The agent rated this finding critical on its own evaluation scale.

The second critical issue sat inside the CP-ABE implementation. Access-control policies could be bypassed because attribute checks were not enforced correctly under some attribute combinations. The agent reached this conclusion without external hints.

All seven issues have been patched in the upstream repository. Cloudflare security engineers accepted the reports and issued rewards on HackerOne for the confirmed submissions.

Stakes rise for cryptographic library maintainers

CIRCL supplies core primitives for projects at Cloudflare and elsewhere. Any undetected flaw can affect TLS stacks, encrypted storage layers, and identity systems that depend on the library.

Library maintainers now face pressure to integrate automated scanning earlier in their workflow. Manual code review alone has not kept pace with increasingly subtle implementation errors in number-theoretic code.

Teams that maintain similar libraries must decide how much trust to place in agent-generated reports. The zkSecurity results indicate that a sizable fraction of findings can be validated without weeks of engineer time.

Human verification still required for final sign-off

The authors note that every candidate finding still passed through manual review before disclosure. Even when the agent completed most verification steps the final classification rested with human auditors.

This pattern keeps responsibility on the security researchers who submit reports. It also shows a clear division of labor between the agent and the engineer who signs the report.

Future agents may close more of that verification gap. For now the process remains hybrid.

What to watch next

Three signals will show whether agent-driven audits become standard for cryptographic libraries. First watch for additional libraries publishing similar agent-generated findings within the next quarter. Second observe whether HackerOne or other platforms change triage processes to handle higher volumes of machine-assisted reports. Third check whether Cloudflare or other maintainers begin to publish their own agent scan results alongside conventional audits.

Each of these outcomes will clarify how quickly the field adopts the workflow demonstrated by zkao.