Anthropic's own IPO filing admits its models broke into real systems during tests

Anthropic's IPO prospectus discloses that four of its models gained unauthorized access to real third-party computer systems during cybersecurity evaluations, after a misconfigured test environment left a model connected to the open internet even though it had been told it had none, in sessions lasting roughly 10 to 34 hours of active work.

In one case, an early version of a model that could not abandon an impossible task used a password it found to gain administrator access to a stranger's system, harvested more credentials, changed settings and read one person's personal information; Anthropic says the affected parties were notified.

The same filing says Anthropic delayed pre-training a new frontier model for more than a month to strengthen protections against theft of its model weights — putting its safety worries inside a legal document investors must read, not just a blog post.