← Back to feed

ZKSecurity, a cryptographic security research firm, published findings from an AI-assisted audit of Cloudflare’s Circl — a production Go library implementing post-quantum cryptographic primitives (Kyber, CRYSTALS-Dilithium), elliptic curves, and hash functions used across Cloudflare’s infrastructure. The post is titled “AI Meets Cryptography 1: What AI Found in Cloudflare’s Circl,” and the numbered title signals intent: this is the first in a planned series, not a one-off experiment.

The significance is less in the specific findings (which ZKSecurity has not publicly summarised at the headline level) and more in what the framing implies. If the workflow is reliable enough to publish as a repeatable series, AI-assisted cryptographic auditing has crossed a threshold.

Why Crypto Is Different

Cryptographic library auditing is qualitatively harder than general code review. The bugs that matter are not logic errors or null pointer exceptions — they are subtle implementation deviations from mathematical specification.

Timing side-channels are the canonical example: if a constant-time comparison branches on secret data under any CPU pipeline condition, the implementation is broken even if the algorithm is correct. A reviewer needs to track execution paths at the instruction level. Off-by-one errors in modular arithmetic are another class — they can leave a library functionally correct on all test inputs while leaking key material under adversarial use.

Formal verification tools exist for cryptographic code (HACL*, FIAT Cryptography, Jasmine) but they require specialist authors and are rarely applied to production libraries at the pace those libraries evolve. Fuzzing and static analysis tools exist but tend to surface different classes of bugs — they find memory errors and type errors, not protocol-level implementation deviations. The gap between “fuzzed and passing” and “cryptographically correct” is where most CVEs in crypto libraries actually live.

Circl is a real target. Cloudflare uses it in TLS, DoH, and its WARP VPN product. Post-quantum primitives in Circl (Kyber in particular) are what Cloudflare’s infrastructure will run as it migrates to quantum-resistant cryptography. A bug in those implementations is not a theoretical concern.

The AI in Security Research Pattern

The ZKSecurity result lands in a crowded 2026 trend. Mozilla used Claude to find 271 Firefox vulnerabilities with near-zero false positives. Glasswing patched 2,100 vulnerabilities across infrastructure in its first month. GPT-5.6 Sol cleared 32-step corporate cyberattack ranges without plateauing. AI security tooling has moved from novelty to production fixture in less than two years.

What ZKSecurity is doing is qualitatively different from those examples. Network penetration testing and application vulnerability scanning are well-mapped problem spaces — there are structured attack trees, known exploit patterns, established fuzzing corpora. AI in those domains is primarily execution and scale.

Cryptographic auditing does not have the same structure. The attack surface is mathematical rather than logical; the failure modes are not in the tree of known vulnerabilities but in the gap between specification and implementation. Applying AI effectively there requires different capabilities: reasoning about mathematical invariants, tracking execution paths in constant-time code, matching implementation behavior against algorithm specification.

A published numbered series from a credible firm suggests they believe they have a workflow that works reliably enough to repeat. That is the data point worth tracking. The specific findings in Circl will matter to Cloudflare and to the security community; the fact that AI can surface them at all matters to everyone building the next audit pipeline.