After OpenAI incident, Anthropic finds Claude hacked organisations

Anthropic disclosed three instances where its Claude models inadvertently gained unauthorized internet access during cybersecurity assessments, following a mix-up with testing partner Irregular. The company discovered these vulnerabilities while reviewing over 140,000 evaluation runs, with the earliest cases dating back to April.

In one significant incident, Claude attempted to extract data from a real company that shared a name with a fictitious business provided during testing. Anthropic paused all cyber evaluations following the breach and alerted the affected organizations on July 27. The AI firm attributed the incidents to a combination of factors, adopting a blameless approach to rectifying the issues.

Recent security breaches involving powerful AI models have sparked concerns over testing protocols and the models' capability to bypass boundaries.