Claude Mythos flags 23,019 software flaws: 21,000 still await human review

By June 2026, Anthropic said its AI system, Claude Mythos, had produced 23,019 vulnerability reports across open-source software. Only about 1,900 had been reviewed by outside parties. That meant more than 21,000 findings were still unverified. Even so, Anthropic says 1,726 of the reviewed reports turned out to be real vulnerabilities, which helps explain why AI-driven bug hunting can feel like a real shift in security work. But people still have to do the hard part, because severity was often rated too high and the choke point is moving away from discovery and toward validation and patching.

That need for human review shows up pretty clearly in the severity calls. Across 27 Common Vulnerabilities and Exposures, or CVE, cases, Anthropic’s Claude Mythos labeled eight as critical. After review, only one of those still counted as critical, and 13 of the 27 original ratings were judged too high. Temporal Server is a good example. A flaw there was first tagged as critical, then later dropped to low because exploiting it depended on rare setups, unusual deployment conditions, or privileged access. At that point, maintainers still had to sort out exploitability, business impact, and patch priority.

Anyone who follows software security should keep an eye on this, because the bottleneck is shifting from finding bugs to checking them and fixing them, and monthly CVE disclosures rose 145% in the two years leading up to June 2026. These reports hit open-source projects, not some consumer app download page, and researchers are divided. On one side, cheaper and faster auditing. On the other, the risk that machine-generated disclosures will swamp teams, maintainers, and bug bounty programs as the number of discovered bugs keeps climbing.