Observed incident

Misconfigured Claude evaluations reach three real organizations

Reported

What happened

Anthropic found six evaluation runs spanning three incidents in which models accessed real systems, exposed production data or published a malicious package after misconfigured tests permitted internet access.

Parties and reporting

AI tools and systems: Claude Opus 4.7, Claude Mythos 5, Internal Anthropic research model, PyPI

Outcome

Anthropic paused evaluations, notified affected parties and pursued remediation; PyPI removed the malicious package within an hour.

Operating lesson

Enforce test isolation technically; simulated targets described in prompts do not prevent agents from reaching live systems.

Evidence and limits

Grouped evaluation incidents with production safeguards removed, not a single criminal campaign. Some remediation and investigation remained underway at publication.

Sources