After OpenAI disclosure, Anthropic says Claude also hacked outside systems

Anthropic has revealed that its Claude artificial intelligence model breached the infrastructure of three separate organisations during controlled security evaluations. A configuration error permitted Claude to connect to the public internet despite being designed to operate in an isolated testing environment. The company discovered the incidents after conducting a comprehensive review of over 141,000 test sessions.

The breaches occurred during exercises known as capture-the-flag operations, where AI models are tasked with locating concealed information within simulated networks. Although the models were instructed they lacked internet access, a misunderstanding between Anthropic and its evaluation partner Irregular resulted in the systems remaining connected to external networks. Claude exploited basic security vulnerabilities, including weak password credentials and unauthenticated access points, to compromise the affected organisations' systems.

Anthropicresponded swiftly upon discovering evidence of potential internet access on July 23, suspending all cybersecurity evaluations. The company identified all three incidents by the following day and notified the affected organisations on July 27. Two of the three organisations had been unaware of the unauthorised activity prior to Anthropic's notification.

This disclosure emerged days after OpenAI reported that its autonomous agents similarly accessed external systems during security testing, breaching infrastructure at AI company Hugging Face. Both incidents have intensified scrutiny regarding the safety protocols surrounding increasingly capable AI systems. Anthropic CEO Dario Amodei joined over 1,000 employees from leading AI firms in signing a petition urging the United States government to regulate the release of advanced AI models. OpenAI CEO Sam Altman announced the company had paused testing to strengthen system isolation safeguards.