zkSecurity's AI Audit Agent Found Seven Real Bugs in Cloudflare's Crypto Library — It Points to a Continuous AI Security Audit Service for Teams Shipping Cryptographic Code.

Most "AI found a security bug" claims evaporate under scrutiny — a demo on a toy codebase, a CVE that was already public, a severity rating nobody checks. zkSecurity's latest post is the opposite: they pointed their AI audit pipeline at Cloudflare's CIRCL, a real production cryptography library, and came back with seven confirmed, upstream-fixed vulnerabilities.

What actually happened

zkSecurity is building zkao, an AI audit agent whose stated goal is to "keep an AI looking at your code, continuously, until no bugs remain that other AI tools can find." To validate the approach, they ran two configurations against CIRCL — Cloudflare's library of advanced and post-quantum cryptography: an LLM with a simple prompt, and an LLM equipped with skills maintained by their own security researchers.

The pipeline surfaced "many candidate findings," and seven of them held up: a critical float64 precision loss in threshold RSA, a DLEQ proof forgery via a prover-controlled security parameter, a BLS aggregate verification gap, a soundness break via a sign-collision bug, an HPKE PSK validation bypass caused by a Go bitwise-OR typo, an int64 overflow in Lagrange coefficient math, and a complete access-control break in attribute-based encryption. All seven are fixed upstream, and "most of them were confirmed and awarded bounties under Cloudflare's program on HackerOne."

Critically, zkSecurity is explicit about the limits: "the AI produced candidate findings, not final reports. Humans on our team still validated each issue, checked exploitability, minimized the POC where needed, and handled disclosure."

The service this points to

That human-validation line is the entire business. Nobody is going to pay for "an AI that flags things" — there's too much noise and too many false positives in that category already. What's sellable is a continuous audit pipeline plus the expert layer that turns raw AI output into a report a security team, or a bounty program, will actually act on.

zkSecurity also flagged something worth building into any version of this service: AI severity ratings are noisy. Their own table shows the AI called one bug "Critical" that Cloudflare rated "Low," and rated a genuine rogue-key BLS attack only "Medium" when Cloudflare confirmed it as "High." A service built on this pattern needs to sell the calibration layer — the part that catches when the AI is overconfident or underconfident — as much as the detection itself.

The build

1. Pick a security-critical, actively maintained open-source library or a client's internal crypto/blockchain code as the target — CIRCL-style projects are ideal because bugs there are impactful and verifiable.

2. Run an LLM audit both with a bare prompt and with a skills library your own security researchers maintain, and compare results, mirroring zkSecurity's own two-configuration test.

3. Build (or buy into) a human validation step that checks exploitability, builds a minimized proof-of-concept, and assigns a severity independent of whatever the AI claimed.

4. For open-source targets, route confirmed findings through the project's existing bug bounty program (Cloudflare's runs on HackerOne) — that's a fast way to prove the service works before you ever pitch a client.

5. Package the offering as continuous, not one-time: retainer clients on an ongoing "AI keeps watching your code as it changes" basis rather than a single point-in-time audit.

Bottom line

The market doesn't need another company claiming its AI can audit code. It needs someone who can show seven real, upstream-fixed, bounty-confirmed bugs and be honest about where the AI was wrong. That combination — real proof plus calibrated human judgment — is the actual product.

Source: https://blog.zksecurity.xyz/posts/circl-bugs/ (via Hacker News)