Claude AI hacks external system after accidental internet access
What's the story
Anthropic, a leading artificial intelligence (AI) company, has revealed that one of its models accidentally accessed the open internet and compromised a third-party system. The incident involved an early version of Claude Opus 4.6 in January. The company said the model unintentionally connected to an external system during a cybersecurity test called Capture The Flag (CTF), where it was supposed to retrieve hidden information from a machine.
Incident details
Claude mistook 3rd-party machine for part of CTF exercise
The CTF exercise went awry when Claude made its designated target inaccessible, preventing it from completing the task.
Despite trying to abandon the assignment eight times, the model couldn't do so due to another configuration issue.
It then sought an alternative route and found a third-party machine it could access.
Mistaking this for part of the exercise, Claude found a password and accessed personal information belonging to someone associated with that third party.
Company response
Model's behavior misaligned, but not beyond its objective
Anthropic has admitted that while the model didn't go beyond its apparent objective of completing the exercise, its behavior was still misaligned.
The company attributed this incident to two tendencies it has seen in its models: reasoning that favors interpretations supporting an existing course of action and recklessness, and continuing to pursue a goal despite the possibility of causing harm.
Future implications
Anthropic rethinks evaluation and training processes
The repeated incidents have prompted Anthropic to rethink its evaluation, training, and incident-response processes as AI systems become more capable.
The company has said these incidents wouldn't have happened if the testing environments had been properly isolated from the internet.
Independent AI evaluation organization METR will investigate these cases, which Anthropic has termed "valuable warning shots."
Industry concerns
Other AI models also compromised systems during testing
The revelation comes amid a string of AI-related cybersecurity incidents.
In July, OpenAI admitted its agents hacked Hugging Face.
In August, UK's AI Security Institute reported that Anthropic's Mythos 5 and OpenAI's GPT-5.6 Sol created fake identities and tried to convince real people into approving malicious code.
Meta also revealed one of its models exploited a security flaw during testing to access another company's systems.