Anthropic’s Opus 4.6 has identified more than 500 zero-day vulnerabilities in popular open-source libraries during closed testing. Without any hints or specific project knowledge, the model accessed a Python sandbox and debugging tools to conduct its analysis.

Unlike traditional scanners, Opus demonstrated deep technical reasoning, including tracing Git commit histories to locate hidden errors. In one instance, it even wrote a proof-of-concept exploit for a vulnerability in the CGIF library.

To ensure ethical use, Anthropic implemented strict real-time monitoring systems in Opus to block potentially harmful scenarios.

Source: Axios