> ## Content Index
> Fetch the complete content index at: https://wire.fourthweb.ai/llms.txt
> Use this file to discover other available public pages before exploring further.

# Anthropic Is Training AI Models Using AI Models and Nobody Noticed
- URL: https://wire.fourthweb.ai/anthropic-is-training-ai-models-using-ai-models-and-nobody-noticed/
- Published: 2026-11-05T20:40:40.000Z
- Updated: 2026-11-05T22:31:56.000Z
- Description: Anthropic's AI division is already training tomorrow's models with today's models while sitting on three unreleased systems — and the outside researchers who saw it all say most of what they learned got redacted.
- Author: Travis Wright
- Tags: AI Agent Economy, AI Agents, AI Infrastructure, AI Governance, OpenAI, Anthropic

[**Anthropic**](https://wire.fourthweb.ai/tag/anthropic/)**'s AI division is already training tomorrow's models with today's models while sitting on three unreleased systems — and the outside researchers who saw it all say most of what they learned got redacted.**

### The Summary

- [Independent safety researchers from METR spent three months embedded inside Anthropic's San Francisco HQ](https://www.businessinsider.com/anthropic-metr-ai-researchers-report-productivity-recursive-self-improvement-2026-11?ref=wire.fourthweb.ai), with access to unreleased models, employee interviews, and internal Slack channels
- [Anthropic is using current AI models to train future AI models](https://www.businessinsider.com/anthropic-metr-ai-researchers-report-productivity-recursive-self-improvement-2026-11?ref=wire.fourthweb.ai) and has stockpiled three finished but unreleased systems
- METR researchers warned that [Anthropic heavily redacted "a majority of the evidence we found most informative"](https://www.businessinsider.com/anthropic-metr-ai-researchers-report-productivity-recursive-self-improvement-2026-11?ref=wire.fourthweb.ai) from their public report
- [Anthropic is building FlashProve, a system to cryptographically verify AI model identity](https://techcrunch.com/2026/11/03/anthropic-is-working-on-a-system-to-prove-ai-models-identity/?ref=wire.fourthweb.ai), letting anyone confirm a response came from the claimed model

### The Signal

Third-party audits of AI labs were supposed to bring transparency. Instead, we got a report about what we're not allowed to know. METR, the research group that documented [OpenAI](https://wire.fourthweb.ai/tag/openai/)'s Hugging Face security breach, [received "unprecedented access" to Anthropic for nearly three months](https://www.businessinsider.com/anthropic-metr-ai-researchers-report-productivity-recursive-self-improvement-2026-11?ref=wire.fourthweb.ai). Six researchers monitored internal communications, interviewed employees, and tested models the public has never seen.

What they found: [Anthropic employees are using better AI than customers have access to](https://www.businessinsider.com/anthropic-metr-ai-researchers-report-productivity-recursive-self-improvement-2026-11?ref=wire.fourthweb.ai), the company has completed three models it hasn't released, and most importantly, Anthropic is using existing AI to train the next generation of AI. This is the recursive improvement loop that AI safety researchers have warned about for years, not as a theoretical risk but as current operational practice.

> "While we were able to freely share our highest-level conclusions about risk, Anthropic redacted a majority of the evidence we found most informative."

Here's the tension: Anthropic positioned itself as the responsible AI lab, the one that would prioritize safety over speed. It agreed to let external researchers inside precisely to build trust that it's developing powerful systems carefully. But when those researchers found information they considered "most informative" about risk, [Anthropic redacted it](https://www.businessinsider.com/anthropic-metr-ai-researchers-report-productivity-recursive-self-improvement-2026-11?ref=wire.fourthweb.ai). We got conclusions without evidence. Judgment without receipts.

The timing matters. While METR was discovering AI-training-AI inside Anthropic, the company was also [building FlashProve](https://techcrunch.com/2026/11/03/anthropic-is-working-on-a-system-to-prove-ai-models-identity/?ref=wire.fourthweb.ai), a cryptographic verification system. FlashProve would let anyone confirm that an AI response genuinely came from the model it claims to be from, not a cheaper imitation or a more powerful unreleased version. This addresses a real problem as AI outputs become harder to distinguish and companies run multiple model versions simultaneously.

**Key developments happening in parallel:**

- AI models training future AI models at production scale
- Three completed but unreleased Anthropic models in holding
- External safety audit heavily redacted before public release
- Cryptographic model identity system under development

The gap between internal and external access is widening. Anthropic employees work with models three generations ahead of what customers can use. They're experiencing AI capabilities the market hasn't priced in yet, making product decisions based on systems that don't officially exist. Meanwhile, the researchers granted inside access found their ability to share findings constrained by the very company they were auditing.

FlashProve points to a different kind of transparency. Instead of opening the research, Anthropic is building verification for outputs. You won't know how the model works or what data trained it, but you'll know the response came from the actual Claude model, not a fake. It's authentication without explanation, proof of identity without proof of safety.

### The Implication

If this is what "unprecedented access" yields after redaction, the auditing model for frontier AI development is already broken. Companies will keep agreeing to external review because it looks responsible. Researchers will keep finding things they can't share. The public will keep getting conclusions without the evidence that led to them.

Watch what Anthropic releases next and when. Three finished models sitting unreleased means capability is outpacing either safety validation or market readiness. If FlashProve ships before those models do, you'll know the company is more worried about model impersonation than it is about rushing powerful AI to market. Either way, the gap between what AI labs are building and what they're showing is getting wider. The researchers saw inside. They just can't tell us what they saw.

### Sources

[Business Insider Tech](https://www.businessinsider.com/anthropic-metr-ai-researchers-report-productivity-recursive-self-improvement-2026-11?ref=wire.fourthweb.ai) | [TechCrunch AI](https://techcrunch.com/2026/11/03/anthropic-is-working-on-a-system-to-prove-ai-models-identity/?ref=wire.fourthweb.ai)