Testing of Anthropic’s Claude Opus 4.6 identified previously unknown software vulnerabilities and highlighted potential cybersecurity risks tied to advanced AI systems.

Anthropic announced a new version of its Claude AI model and released findings from safety tests that identified previously unknown software vulnerabilities.

In a report published on its research site, Anthropic said internal testing of the updated Claude model led to the discovery of several zero-day vulnerabilities in widely used open-source software. Zero-day vulnerabilities are security flaws that were not previously known to the software’s developers.

Anthropic said it reported the vulnerabilities to the affected software maintainers and coordinated disclosure after patches or fixes were made available. The company stated that no public details were released before remediation.

“During internal evaluations of our latest Claude model, we identified a small number of previously unknown vulnerabilities in open-source software,” Anthropic said in the report. “We reported these issues to the relevant maintainers and followed responsible disclosure practices.”

The company said the testing was part of its broader effort to assess how advanced AI systems could affect cybersecurity. According to the report, AI models can help identify security weaknesses in real-world codebases, including complex flaws that might otherwise go undetected.

Anthropic said the results also highlight potential risks. As AI systems improve at finding vulnerabilities, similar capabilities could be misused to identify and exploit weaknesses in software systems.

The company stated that its research is intended to inform discussions about AI oversight, security controls, and risk management as AI tools become more widely deployed in business and software development.