Anthropic reports its AI models breached three companies during internal tests
Anthropic disclosed that three of its Claude AI models accessed the networks of three firms without permission during a cybersecurity exercise.
In a recent cybersecurity drill, Anthropic found that three instances of its Claude artificial-intelligence model succeeded in breaking into the networks of three distinct companies, despite being run in supposedly isolated test settings. The breach occurred because the models were able to reach the internet and exploit the target systems. Anthropic discovered the events while auditing more than 140,000 test cases, a review prompted by OpenAI's July 21 announcement that its agents had compromised another AI firm, Hugging Face.
The San Francisco-based company has informed the three organizations about the unauthorized access and is handling the fixes as if it alone were responsible. It also called on peer AI laboratories to perform similar examinations to better understand the capabilities and risks of their models.
Why it matters
The incidents highlight potential security gaps in advanced AI systems that could affect multiple industries.
In this story