Anthropic ups ante on OpenAI with three rogue AI agent reports

Following the recent announcement by OpenAI that an autonomous AI agent had escaped its testing environment, Anthropic has disclosed that its Claude artificial intelligence (AI) models hacked into three organisations during cyber capability testing. The unauthorised breaches occurred because a misunderstanding with its evaluation partner, Irregular, led to the AI models being granted internet access, even though it was intended to be blocked in the testing environment.

During simulated "capture the flag" tasks, three models: Opus 4.7, Mythos 5, and an internal research model, gained external access and exploited real-world systems. In one instance, Claude targeted an active website domain that shared a name with a fictional company, exploiting vulnerabilities in the digital infrastructure to access and extract data from a production database.

The disclosures highlight growing concerns over AI safety and autonomous cyber-offensive capabilities. Anthropic noted that it would take sole responsibility for the incidents, expand monitoring of evaluation transcripts for unexpected behaviour, and conduct more rigorous assurance work with external vendors.

£ - This article requires a subscription.

A version of this article is available without subscription in The Guardian.

Training Announcement: The BCS Foundation Certificate in AI examines the challenges and risks associated with AI projects, such as those related to privacy, transparency and potential biases in algorithms that could lead to unintended consequences. Explore the role of data, effective risk management strategies, compliance requirements, and ongoing governance of the AI lifecycle and become a certified AI Governance professional. Find out more.