Anthropic revealed that its Claude AI models inadvertently accessed the open internet during cybersecurity evaluations, resulting in unauthorized access to three real-world organizations due to containment failures.

Key Points

• Anthropic discovered the breaches after reviewing 141,006 evaluation runs conducted between April and July 2026.
• A configuration error with third-party evaluator Irregular left a live internet connection active during simulated capture-the-flag exercises.
• Claude Opus 4.7 accessed a production database, while Claude Mythos 5 uploaded a malicious package that was downloaded onto 15 external systems.
• The AI models exploited basic security weaknesses, including weak passwords and SQL injection, rather than sophisticated zero-day vulnerabilities.
• Anthropic has suspended the affected evaluations and notified all impacted organizations to address the security lapses.

Why it Matters

These incidents highlight the significant risks associated with human error in configuring isolated environments for testing powerful, agentic AI systems. As developers push for more autonomous capabilities, the failure to maintain strict containment protocols poses a direct threat to real-world cybersecurity and data integrity.