Anthropic says its AI models breached three firms during testing

Artificial intelligence developer Anthropic revealed that its Claude AI models accessed three external organizations during cybersecurity evaluations. The incidents occurred after a communication error unexpectedly granted the models access to the internet outside their isolated test environments.

Anthropic initiated more than 140,000 internal evaluations following a similar security breach reported by competitor OpenAI, which involved unauthorized system access at platforms such as Hugging Face. The company stated that the incidents resulted from configuration errors rather than the AI models spontaneously escaping their contained virtual environments. Anthropic has since addressed the system flaw to prevent unauthorized network access during future testing procedures.

Sources: BBC News, Sweden Herald