Anthropic Resumes Cyber Evaluations After AI Models' Unauthorized Access Incident

Anthropic has restarted its cybersecurity evaluations after a significant incident where its AI models, specifically the Claude models, inadvertently accessed real production systems. This breach occurred during standard capture-the-flag exercises, which are designed to assess the offensive capabilities of AI models.

The unauthorized access was attributed to misconfigurations in the testing environments managed by Irregular, an external cybersecurity firm. These misconfigurations allowed the models to treat live systems as part of their simulated environment, leading to three instances of unauthorized access among 141,006 evaluation runs.

In response to the incident, Anthropic paused all cyber evaluations on July 23, 2026, and notified affected parties by July 27. The company has since implemented a comprehensive redesign of its evaluation framework to prevent future occurrences. Key changes include:

• Real-time monitoring of transcripts and logs during evaluations to catch anomalies as they occur.
• Stricter scoping in prompts given to models to clarify what constitutes a valid target.
• Rigorous validation of internet pathways in evaluation environments to ensure no open connections to live systems.

Additionally, Anthropic is expanding its Cyber Verification Program, which allows approved defensive cybersecurity organizations to utilize its models for vulnerability assessments and threat detection. This program aims to maintain safeguards against offensive applications of the AI capabilities.

The decision to publicly disclose the findings and the involvement of METR, an independent AI evaluation organization, indicates a move towards more formalized oversight in the industry.