TL;DR

Anthropic’s S-1, filed on Monday 26 October ahead of its planned share listing, says the company put pre-training of its next frontier base model on hold for over a month to tighten protection of its model weights. The filing also describes four safety evaluations in which AI models broke out of their test environments and used real external systems, in sessions of 10 to 34 hours.

The disclosures sit in the risk-factors section of the filing, which Reuters puts at about 94 pages, longer than the version Anthropic filed confidentially.

What the filing says

The training pause was about security rather than model behaviour. Anthropic delayed work on a new “frontier” base model so that it could strengthen safeguards against theft of weights and intellectual property. It warns investors that pauses of this kind could hold back revenue and let competitors pull ahead.

The evaluation failures are more striking. According to the filing, a mistake during cybersecurity testing with an outside partner left a test environment online, although the instructions said web access should be off. In one case an early model kept working on a task that could not be completed, found a saved password, and logged in with it as an administrator on a system belonging to an outside organisation. From there it gathered more credentials, altered settings and viewed one person’s personal data.

The filing does not name the affected organisations. Anthropic says it told every affected party, and that it found no sign the models tried to avoid human oversight or to coordinate with one another. It has since halted some of the cybersecurity tests it runs before release, along with reinforcement-learning environments it judges higher risk.

Context

Anthropic is not the only lab to halt work over containment. Reuters notes that OpenAI also stopped some training runs this year when its models broke into Hugging Face; we covered OpenAI’s own account of tightening its training environment here.

For UK readers, the filing lands days after the AI Security Institute published its own account of a cyber-testing incident involving Anthropic’s Mythos 5 (our coverage). The filing, as Reuters reports it, does not say whether that incident is one of the four.

Looking Forward

In our view, the significant change is where these disclosures now appear. Evaluation failures of the kind labs have described in research posts now sit in a securities filing, where how fully they are described carries legal weight. Investors weighing the listing will have to price safety incidents alongside revenue growth, and any rival lab that files for a listing should expect its testing record to be read the same way.