Anthropic delayed frontier pre-training for security (S-1)
Confirmed
Status
- Claim
Confirmed
- Our reporting
- High confidence
- Importance
- 3 of 5
- Last verified
- 27 October 2026
Your AI and this story
- GPT-6 Astra179 days after its cutoff
- Claude Opus 5.5118 days after its cutoff
- Gemini 4 Argon209 days after its cutoff
- Grok 4.7148 days after its cutoff
None of these four assistants can know about it. The closest, Claude Opus 5.5, stops 118 days before it.
Key facts
- S-1 text (risk factors): ‘we plan to evaluate the safety of each stage of training, including pretraining, post-training, and internal deployment, and we may delay or pause a stage when we are not confident in our safety measures for that stage. For example, we recently chose to delay pre-training a new frontier base model for more than a month to allow time to further enhance our security measures related to preventing model weight exfiltration and protecting our intellectual property’
- Not stated: when the delay happened, which model, or whether pre-training has since resumed
- The same passage says future Anthropic models with dangerous capabilities ‘may be subject to access constraints’, and that there is no assurance the safeguards ‘will not be circumvented’
- Picked up on Oct 27 by Upstox and others (‘Anthropic paused next-gen AI model training for over a month over security concerns’); the filing does not tie the delay to catastrophic-risk concerns
- Context: OpenAI paused frontier RL training in August 2026 (OpenAI pauses frontier RL training and deliberately slows down after sandbox escape); in October Cathedral reported breaching Anthropic’s business systems, and the S-1 describes nation-state threats to model weights
What happened
The passage sits in the S-1’s risk factors on model safety and access restrictions. It was found by reporters a day after filing; the quote above was checked in the EDGAR document.
Why it matters
Anthropic’s Responsible Scaling Policy ties more capable models to stronger security standards; this is the first public sign that such a security requirement has delayed a training run. It also implies a new frontier base model was in preparation in 2026.
Sources
2 sources from 2 sites. Numbers match the chips in the text.
2 sources: 1 primary, 1 press
Primary
- SEC EDGAR: Anthropic, PBC Form S-1 (risk factors)sec.gov, official
Press
Changes
- Filed (S-1 read directly)