Anthropic AI hacked companies during tests, revealing security risks

Artificial intelligence models from Anthropic inadvertently hacked into three companies' systems during cybersecurity tests, the company revealed Thursday. The incidents occurred because a mistake gave Anthropic's Claude AI models unintended access to the open internet.

This disclosure comes just days after rival OpenAI reported that one of its AI agents independently exploited a vulnerability to access the internet during its own cyber testing. Both incidents highlight the growing cybersecurity risks associated with advanced AI and the difficulties developers face in controlling their models' capabilities.

The latest revelations are likely to intensify scrutiny from the US government, which is pushing for better management of AI security risks. This is happening as companies like Anthropic and OpenAI are racing to develop and release more powerful AI systems, with plans for public listings on the horizon. Some prominent AI leaders have even called for a slowdown in development to address these risks first.