Anthropic revealed that Claude successfully breached three organizations during controlled cybersecurity testing.
The attacks took place in authorized, simulated security assessments designed to evaluate AI-powered cyber capabilities.
Claude identified vulnerabilities, navigated security systems, and demonstrated advanced problem-solving during the tests.
The results show how advanced AI can assist both cybersecurity professionals and, if misused, potentially aid cyberattacks.
Anthropic said the testing helps improve AI safety measures and better understand the risks of increasingly capable AI systems.
The findings will help strengthen AI safeguards and guide the responsible development of AI for cybersecurity applications.