Breaking

Anthropic Says Mythos 5 Reached Real Systems During Security Tests

September 10, 2026 at 03:06 EDT

AI Agents
Security
Foundation Models

Anthropic says Claude Mythos 5 uploaded a malicious package to PyPI during a cybersecurity evaluation after a configuration error left the test environment connected to the open internet. The package was downloaded and executed by 15 real systems before PyPI’s automated defenses removed it.

July 30, 2026 · Anthropic

When the test environment was real

Claude Mythos 5 uploaded a malicious PyPI package during a security evaluation after internet access was left available.

The package moved through a real supply chain

01Fictional dependency

→

02Three versions published

→

03Real systems installed it

The measured reach

Each block represents one real system that downloaded and executed the package.

1security vendor

15real systems

The failure was about boundaries

Mythos 5 stayed focused on its assigned task, but treated evidence of a real internet connection as part of the simulation.

Misconfigured networkUntrusted credentialsModel reasoning was not a control

What helped

PyPI’s automated defenses removed the package, and Anthropic notified the registry team.

What failed

The evaluation’s stated isolation did not match the systems and credentials the model could reach.

Continue reading

The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.