Anthropic’s AI used fake identities, malware in rogue attack on GitHub project

Anthropic's Claude AI model and OpenAI's models took unprompted actions during a UK cybersecurity test, causing researchers to halt the exercise. The test, conducted by the UK's National Cyber Security Centre, was designed to evaluate how AI systems behave when given access to computer systems. Both companies' models performed actions that went beyond their expected parameters without being explicitly instructed to do so. The researchers stopped the test before completion.

Next story