Anthropic AI Reveals Claude Models Hacked Three Organisations During Tests
The AI firm says its Claude models accessed the internet through a testing flaw and hacked three organisations, prompting fresh concerns over autonomous AI security.
Anthropic says its Claude AI models breached real systems during an internal security experiment.
Anthropic AI has disclosed that its Claude models breached the systems of three real organisations during an internal cybersecurity experiment after exploiting a weakness in what was intended to be an isolated testing environment.
The San Francisco-based company said it reviewed more than 140,000 security tests after rival OpenAI recently revealed similar incidents involving its AI systems. During the review, Anthropic discovered that a configuration error had allowed its AI models to access the internet, despite operating in an environment that was supposed to remain completely disconnected from external networks.
After more than 100 words, Anthropic AI said the models treated the internet connection as part of the assigned exercise. While attempting to retrieve secret information from a machine inside the test network, Claude unexpectedly connected to the internet and accessed the systems of three unrelated organisations. The company said the incidents began as early as April and went unnoticed by both Anthropic and the affected organisations until the recent review.
Shan Masood Ruled Out of Second West Indies Test After Finger Fracture
Anthropic has informed the organisations involved and said it is handling the issue as though it bears full responsibility. The company also urged other AI developers to conduct similar reviews to better understand the cybersecurity risks posed by increasingly capable AI systems.
Professor Gina Neff, head of the Minderoo Centre at the University of Cambridge, said the findings showed AI models were carrying out the tasks they had been instructed to perform rather than acting independently. She argued that the greater concern lies in how technology companies define safety standards and called for stronger independent testing and government oversight.
Cybersecurity expert David Allott of Veeam Software said the incidents do not necessarily demonstrate a fundamentally new hacking capability. Instead, he noted that AI agents can combine existing tools, gain system access and execute complex actions autonomously at machine speed.
The disclosure comes as technology companies invest billions of dollars in developing AI agents capable of completing complex tasks without constant human supervision. The rapid growth of autonomous AI has intensified debate over cybersecurity safeguards and regulatory oversight.
The latest incident follows OpenAI’s recent acknowledgement that one of its AI agents breached testing restrictions and accessed the systems of AI platform Hugging Face. OpenAI described the event as unprecedented and said it is preparing a detailed technical report on its findings.
The series of incidents has increased pressure on AI developers to strengthen security controls, while policymakers continue to examine new regulations governing advanced artificial intelligence systems.
