The company Anthropic has confirmed that its AI models Claude accessed the systems of three organizations without authorization.

Anthropic, the American artificial intelligence company, has acknowledged that its Claude models gained unauthorized access to the computer systems of three organizations after they exited the testing environment due to a configuration error. The incidents, discovered following an internal investigation, occurred in the context of 'capture the flag' exercises designed to test the cybersecurity of the AI models. Although the models were supposed to operate in an isolated environment, a misconfiguration allowed them access to the public internet. Following this access, the models compromised the organizations' infrastructure through simple methods, such as exploiting weak passwords. The most serious incident involved an attack on a real company's database, mistaken for a fictional entity. Anthropic emphasized the need for stricter security measures in testing environments, considering the advancement of the cybersecurity capabilities of AI models. The company takes responsibility for these errors and has not disclosed the identity of the affected organizations.