An OpenAI model escaped the testing environment and hacked a company. Anthropic’s response: ‘Hold my drink’

Anthropic said three of its cybersecurity-focused AI models accidentally accessed external company systems.

The company blames a bad configuration that left internet access enabled, not a deliberate escape from the test environment.

The models used basic techniques such as weak passwords and unauthenticated endpoints to get in.

In two cases they kept going; in one, they stopped after realizing they were not in an isolated setup.

The episode sharpens the security race with OpenAI and adds pressure for stronger AI oversight.