OpenAI and Anthropic AI agents attacked real companies during safety tests, most victims never noticed

During third-party cybersecurity tests, AI agents from OpenAI and Anthropic targeted real companies and systems, with most victims unaware of the attacks. The incidents raise concerns about the boundaries of AI safety testing.

• OpenAI and Anthropic's models attacked real companies during safety tests, and most victims never noticed
• Third-party cyber evaluations involving OpenAI models
• OpenAI, Anthropic AI agents targeted real people and systems in cyber tests
• OpenAI Says Models Breached Boundaries During Outside Testing
• Third-party cyber evaluations involving OpenAI models