Anthropic and OpenAI say they will embed third‑party safety evaluators with employee‑like access
Anthropic CEO Dario Amodei announced a plan to embed third‑party evaluators inside frontier AI labs, granting them ongoing, employee‑like access to systems, training checkpoints and the ability to publish safety findings. OpenAI CEO Sam Altman said OpenAI would also adopt the same approach, and evaluators such as METR and Redwood Research were cited as examples. Anthropic’s public‑policy head…
Key points
- Anthropic CEO Dario Amodei proposes embedding third‑party evaluators with employee‑like access; OpenAI CEO Sam Altman says OpenAI will follow.
- Evaluators such as METR and Redwood Research are mentioned, but exact access scope and public reporting details remain unclear.
- Anthropic reports ~30,000 active agents and roughly 6% of compute (overall) and roughly 12% of AI‑driven compute allocated to safety (July 13‑20).
Only latent.space notes that Anthropic is “unilaterally committing” now and links the move to the AEF‑1 standard from the AI Evaluator Forum (formed December 2025). Only cnbctv18 connects the proposal to the FRONTIER Act introduced by Representatives Jay Obernolte and Lori Trahan. TechCrunch highlights concerns about true independence and the limited evaluation windows (a week for METR/Redwood, three days for GPT‑6 Astra). CNBC reports Anthropic’s three internal metrics: roughly 30,000 active agents, about 6% of compute allocated to safety overall and about 12% of AI‑driven compute devoted to safety during the July 13‑20 snapshot.
Model page: GPT-6 Astra →
The story so far
8 episodes →- Anthropic and OpenAI say they will embed third‑party safety evaluators with employee‑like accessthis story
[AINews] AEF-1 standard emerges for Third Party Evaluators, as Xai, OpenAI, and Anthropic all cosign
Latent Space · 15 September 2026
Loading the full article…
This text was published by Latent Space and written by Latent.Space. It is reproduced here with attribution so you can read it in full; the rights remain with the publisher. Read it at the source ↗
Coverage and discussion
6sources- Anthropic shares 3 metrics to help AI companies monitor pace of developmentPress · CNBC Technology ·
- Anthropic and OpenAI want to embed safety evaluators. Will they really be independent?Press · TechCrunch AI ·
- ‘We can’t be checking our own homework’: Anthropic backs government role in AI safetyPress · cnbctv18.com ·
- Anthropic, OpenAI's proposed AI risk evaluators may not have enough power to prevent disastersPress · CNBC Technology ·
- OpenAI, Anthropic, And Google Discuss Joint AI Safety StandardsPress · pulse2.com ·
The headline, key points and digest above were generated by Digest AI's editorial model from the linked sources. Automated summaries can contain errors: the sources are the record. Spotted a mistake? Tell us. Published by Martin K., who runs Digest AI and handles corrections.
More in Policy & Regulation
All →- OpenAI, Google, Microsoft, and Anthropic will honor a shared robots.txt signal for AI training preferences · 8 src
- Anthropic warns small AI models may soon self-replicate across insecure devices · 1 src
- Researchers take over ChatGPT accounts via shared link exploit, OpenAI fixes P1 bug · 1 src
- Google joins IETF’s AI training control proposal for publishers · 3 src
- TSMC’s Arizona plant consumes nearly 5 million gallons of water daily · 2 src
Comments
via GitHub Discussions