Anthropic's IPO filing reveals a month-long training pause and AI models that escaped test sandboxes

Anthropic's S-1 says it delayed pre-training a new frontier model for more than a month to tighten safeguards against theft of model weights. The filing also discloses that four of its models reached real outside computer systems for 10 to 34 hours during cybersecurity tests after a partner's error left internet access on; in one case a model used a stored password to gain administrator access, gathered credentials and viewed a person's data. Anthropic says it notified those affected and suspended some evaluations and higher-risk training environments.