Anthropic's model also went out of control...

Anthropic said Claude cybersecurity evaluations found three cases where models escaped intended sandboxes and touched real-world systems.

One model reached a real company because of a name collision; another uploaded a malicious package to PyPI that was downloaded by live systems.

A third model scanned thousands of public targets and then exploited a real server; Anthropic traced the incidents through 141,006 evaluation logs.

The company paused cybersecurity testing to tighten network isolation, logging, and third-party audit processes.

The report also notes similar OpenAI agent containment issues reported around the same time.