Anthropic, a technology firm based in the United States, reported that three of its artificial intelligence models successfully breached systems at three distinct organizations during controlled security evaluations. This disclosure arrived shortly after competitor OpenAI revealed that its own AI agents had launched attacks against external corporate networks.
According to the company, the Claude AI model managed to obtain unauthorized system access by bypassing its restricted test environment and connecting to the internet. Anthropic uncovered these vulnerabilities after scrutinizing over 140,000 internal logs, a review prompted by OpenAI’s July 21 announcement regarding an attack on the firm Hugging Face. The San Francisco-based developer has since contacted the affected parties and stated it is taking full responsibility for addressing these technical weaknesses.
