Anthropic recently dropped a update on its security research. The company revealed that its specialized AI model, Claude Mythos Preview, found more than 23,000 security flaws across roughly 1,000 open-source software projects.
While this is a massive win for AI security, it has triggered an industry-wide crisis: AI is finding bugs faster than humans can fix them. These automated discoveries is completely overwhelming software developers and breaking the traditional system used to patch vulnerabilities.
Key Points: The Mythos Open-Source Scan
In an extensive scan of over 1,000 foundational open-source projects—which underpin much of the modern internet and Anthropic’s own infrastructure—the Mythos model flagged 23,019 potential flaws.
- Severe Threats: Out of those findings, Mythos estimated that 6,202 are high- or critical-severity vulnerabilities.
- High Accuracy (True Positive Rate): Independent security research firms have begun triaging the results to weed out AI “hallucinations.” Of a subset of 1,752 critical flags carefully reviewed by humans, 90.6% were confirmed as valid, true-positive flaws, with 62.4% validated as genuinely high or critical risks.
- Autonomously Generated Exploits: To prove the severity of the findings, Mythos autonomously discovered and constructed a fully functional exploit for a certificate-forgery vulnerability in wolfSSL (CVE-2026-5194, CVSS 9.1), a cryptography library used by billions of devices. The exploit would allow an attacker to host a fake bank or email website completely indistinguishable to an end-user.
- The Patching Bottleneck: Despite thousands of confirmed flaws, fewer than 100 patches have been publicly deployed so far. Open-source maintainers are reportedly “swamped,” with several explicitly asking Anthropic to slow down its disclosure rate because human engineering teams lack the capacity to triage and patch at AI speed.
What is Anthropic’s Project Glasswing?
Project Glasswing is Anthropic’s controlled, collaborative initiative designed to secure systemically important digital infrastructure before next-generation, highly capable AI models are turned against it by malicious threat actors.
Because an AI model capable of finding thousands of bugs for defenders could just as easily be used by cybercriminals to find zero-days to exploit, Anthropic has severely restricted public access to Mythos.
Instead, they launched Project Glasswing to safely distribute the tool to a closed coalition of roughly 50 handpicked partner organizations—including major tech companies, independent security firms, and allied government agencies (such as the UK’s AI Security Institute).
Initial Results from Glasswing Partners:
The update shared by Project Glasswing participants mirrors the overwhelming success seen in the open-source scan:
10,000+ Critical Flaws: Collectively, Glasswing partners have utilized Mythos Preview to discover more than 10,000 high- or critical-severity vulnerabilities across their internal systems in just one month.
- Cloudflare: Uncovered roughly 2,000 software bugs (400 high/critical) across its critical-path systems. Cloudflare’s security team noted that the AI’s false-positive rate was actually lower than that of human penetration testers.
- Mozilla: Used the model to audit its browser codebase, successfully identifying and fixing 271 vulnerabilities in Firefox 150. This represented a massive 10x increase in effectiveness compared to audits done with Anthropic’s older model, Claude Opus 4.6.
- UK AI Security Institute: Reported that Mythos Preview is the first AI model to successfully complete both of its advanced cyber range simulations from end to end.
What’s Next for AI Security?
Anthropic eventually wants to release these powerful “Mythos-class” tools to the general public so everyone can secure their code. However, there is a major catch: nobody has figured out how to stop hackers from using these exact same tools to find and exploit bugs.
Because of this risk, Anthropic is keeping the technology under lock and key. For now, it will only expand secretly through Project Glasswing—a private network of trusted security teams. In the meantime, the tech world faces an urgent challenge: figuring out how human developers can reinvent the software patching process to keep up with the relentless speed of artificial intelligence.
