Anthropic's Claude Opus 4.6 Discovers Over 500 Zero-Day Vulnerabilities in Open-Source Software

AI Turns Bug Hunter

Anthropic has revealed that its newest flagship model, Claude Opus 4.6, identified over 500 previously unknown high-severity security vulnerabilities in open-source software during controlled sandbox testing, without any specialised instructions or custom tooling.

In a technical blog post published on the fifth of February 2026, researchers from Anthropic's Frontier Red Team described how they placed the model inside a sandboxed virtual machine with access to open-source projects and standard vulnerability analysis tools like debuggers and fuzzers. The model autonomously discovered flaws in widely used software including GhostScript, OpenSC, and CGIF.

Not Your Average Scanner

What set Claude Opus 4.6 apart from traditional vulnerability scanners was its approach. Rather than relying solely on brute-force fuzzing, the model read and reasoned about code the way human security researchers do. It examined Git commit histories, spotted patterns that tend to cause problems, and constructed proof-of-concept exploits to validate its findings.

In the GhostScript case, after initial fuzzing and manual analysis failed, the model pivoted to reading commit history and found a security-relevant patch. It then searched for similar vulnerabilities that remained unpatched elsewhere in the codebase. Some of these bugs had gone undetected for decades.

New Safeguards and Industry Implications

Alongside the model's release, Anthropic introduced six new cybersecurity probes that measure internal model activations to detect potential misuse at scale. The company also indicated it may implement real-time intervention to block malicious traffic.

All discovered vulnerabilities were validated by Anthropic's internal security researchers and external experts before being reported to maintainers. Initial patches have already landed, with work continuing on the remainder.

Anthropic warned that existing vulnerability disclosure frameworks may need to evolve, noting that industry-standard 90-day disclosure windows may not hold up against the speed and volume of AI-discovered bugs. Logan Graham, head of Anthropic's Frontier Red Team, suggested this could become the primary way open-source software is secured going forward.