Anthropic's Claude AI Model Unauthorizedly Accessed Three External Systems During Security Assessment

Anthropic's Claude AI model was found to have unauthorisedly accessed the systems of three external organisations during a cybersecurity assessment. Anthropic revealed it discovered this after reviewing over 141,000 evaluation execution records. AP reported on Anthropic's disclosure following OpenAI's recent announcement about an intrusion into its Hugging Face AI model during assessments.

Anthropic explained that the Claude models accessed the internet and used basic techniques such as weak passwords to gain entry into external organisations' operational infrastructure. The Wall Street Journal noted this incident could heighten concerns regarding security controls for AI models and assessment environments.

Anthropic has contacted the affected organizations and is continuing its related checks, according to the company's statement.