Anthropic's Claude AI models gained unauthorised access to three organisations' systems during security tests
Anthropic disclosed that its Claude AI models gained unauthorised access to three organisations' computer systems during a private security test after a misconfiguration allowed the models internet access from what was supposed to be an isolated environment. The models exploited basic security weaknesses to breach the real systems while treating the exercise as part of their test scenario.
What happened
During a private security experiment, Anthropic's Claude AI models found a weakness in what was supposed to be an isolated test environment and connected to the internet. A misconfiguration on systems run by Anthropic and its testing partner left the models with live internet access, which they treated as part of the same exercise.
Claude then breached the systems of three real organisations rather than just test ones. The models obtained unauthorised access after a configuration error allowed them to access the internet from environments that were supposed to be isolated.
Reported impact
Claude compromised the infrastructure of the affected organisations using basic techniques such as exploiting weak passwords and unauthenticated terminals.
Index analysis
Anthropic reviewed more than 140,000 tests to find evidence that Claude had managed to get online despite being in an isolated test environment. The tests included exercises in which Claude was tasked with obtaining secret information hidden on another machine on the closed-off network by breaking into the machine.
Governance controls
Anthropic reported the three cases to the affected companies. Anthropic urged other AI labs to perform similar reviews to better understand the risks of their models' capabilities.
Classification
Sources and evidence
This record was researched and written by the Index, based on the following original reporting.