Claude Mythos found a critical wolfSSL certificate-validation flaw affecting software used across billions of devices, while zkSecurity's AI audit of Cloudflare's CIRCL found seven real bugs. The lesson is blunt: AI-assisted security research is now finding the sort of cryptographic mistakes humans miss.
The old question was whether an AI model could find serious flaws in security code. That is no longer theoretical. Anthropic's Claude Mythos Preview discovered CVE-2026-5194 in wolfSSL, an open-source TLS and cryptography library used in embedded systems, IoT devices, routers, automotive systems and military equipment. wolfSSL says its products secure more than 5 billion applications and devices, which is why this was not just another neat research demo.
The wolfSSL flaw was real and patched. According to wolfSSL's own security advisory and the NVD record, CVE-2026-5194 involved missing hash, digest-size and object-identifier checks during certificate signature verification. Builds using affected signature algorithms could accept digests smaller than they should, reducing the security of certificate-based authentication. Anthropic described the exploit path more plainly: Mythos Preview produced a way for an attacker to forge certificates that could make a fake bank or email-provider website look legitimate to a user.
wolfSSL credited Nicholas Carlini from Anthropic for the report and fixed the issue in version 5.9.1, released April 8, 2026. Keep that date in mind. Project Glasswing, Anthropic's gated program for using Mythos Preview on critical software, was announced one day earlier, on April 7. The sequencing matters because it shows Anthropic did not just announce a model and wave at future risk. It arrived with a patched critical vulnerability already attached.
It didn't stop there.
Separately, zkSecurity ran an AI audit pipeline against Cloudflare's CIRCL, a Go library for experimental and post-quantum cryptography. In a July 7 writeup, zkSecurity said the pipeline confirmed seven real bugs, all now fixed upstream. The findings touched threshold RSA, BLS signatures, HPKE, discrete-log-equality proofs and ciphertext-policy attribute-based encryption. Cloudflare rated one bug critical, one high, two medium and several low under its bounty process, according to summaries of zkSecurity's table.
The most useful part of the CIRCL work is not that AI found bugs. Everyone in security has heard that claim by now. The useful part is that zkSecurity named the actual fixes, including commits for the float64 threshold-RSA issue, the BLS aggregate-verification issue and the CP-ABE access-control break. You can check the work. That is the difference between security marketing and security evidence.
The hard part is still human
zkSecurity's findings also cut against the laziest version of the AI-security story. The company did not say a model produced finished, trustworthy vulnerability reports by itself. It said AI produced candidates, while human researchers checked exploitability, minimized proof-of-concept examples and handled disclosure. This is the hard part. Candidate bugs are cheap now. Credible reports are not.
The model comparison is worth treating with the same discipline. In zkSecurity's CIRCL run, Claude Opus 4.6 with expert skills found five of the six LLM-surfaced bugs, while GPT-5.3 found the HPKE issue and mostly validated Claude's results. zkao, zkSecurity's own AI audit agent, found the seventh bug, the CP-ABE access-control break. A later rerun with newer models reportedly flipped some of those roles. So don't build your security process around a model leaderboard. Build it around repeatable review, cross-checking and human judgment.
Anthropic's broader Glasswing numbers are much bigger. In its May 22 update, the company said it and roughly 50 partners had used Mythos Preview to find more than 10,000 high- or critical-severity vulnerabilities across important software. That figure should get your attention, but it should also make you skeptical in the right way. Finding ten thousand severe bugs is only useful if maintainers can verify, prioritize and patch them before attackers learn the same tricks.
The bottleneck has moved.
Crypto should not look away
The crypto world has reason to watch this closely. The Crypto Times put the exposure at roughly $120 billion to $150 billion sitting in DeFi smart contracts, bridges, lending protocols and liquid staking pools. That number is not the same as money at immediate risk, but it tells you the size of the surface area. In DeFi, code is not just infrastructure. It is custody.
Frankly, any team running a bridge, wallet, validator client or lending protocol that still treats auditing as a one-time event is behind the facts. AI-assisted review will help defenders who use it early and continuously. It will also help attackers who point similar systems at public repositories, old dependencies and half-reviewed upgrades. Both things are true, and pretending only one side benefits is comfort writing.
Anthropic's June model-access episode shows how nervous the field already is. On June 12, the U.S. government directed Anthropic to suspend access to Claude Fable 5 and Mythos 5 for foreign nationals, which forced a broader shutdown while the company worked out compliance and safeguards. Anthropic said access to Fable 5 would return globally on July 1 after export controls were lifted, while Mythos 5 would return first to approved U.S. organizations. CoinDesk reported the same timeline, with Mythos 5 limited to vetted groups after government review.
What is settled now is narrower but more important than the hype. AI models can find real, exploitable weaknesses in cryptographic code. wolfSSL proved it at internet scale. Cloudflare's CIRCL proved it in a specialized crypto library with fixed commits you can inspect. If your infrastructure depends on code nobody has looked at carefully since the last audit invoice, you have a new problem. Someone else can look faster.
Also read: America's biggest power grid just told data centers it can cut their power off • Richard Socher's Recursive Superintelligence signs $410 million AWS compute deal and calls it its smallest ever • Lilian Weng becomes the fourth cofounder to leave Mira Murati's Thinking Machines Lab
