Anthropic gave six METR researchers employee-level access and no veto over what they publish. Musk says SpaceX will probably follow
Nine days after Reuters reported that labs were negotiating editorial control over the evaluators embedded inside them, Anthropic published the terms of its own arrangement. Six METR researchers now have company laptops, office space and internal tools roughly on par with Anthropic's own risk teams. METR publishes what it finds and Anthropic has no veto. It can require redactions in narrow, predefined categories, Anthropic decides what falls inside those categories, and METR is free to say publicly whether it thought a redaction was justified. The post, dated October 15, frames it with a bank-examiner analogy and ends with the line that every frontier AI company should do the same. TechCrunch's framing is the useful one: this landed after rogue-agent incidents arriving every week or two since late August.
The structure is a three-tier system, and the tiers are the news. Tier one is narrowly scoped work where Anthropic approves publication. Tier two is broader access where the evaluator publishes without approval. Tier three is the METR arrangement. SecureBio is reviewing the chemical and biological weapons chapter of the August risk report in the bottom tier. Accenture and Faculty have a partnership, and Anthropic says it is in early discussions with Hugging Face. The design answers the two questions the Reuters story raised: who decides what gets published, and whether the evaluator sees enough to matter. The redaction categories are the remaining lever, and Anthropic kept it.
The reaction was fast and a little strange. Elon Musk posted on October 16 that SpaceX will probably allow external evaluators as well, following Anthropic, which The Verge summarized as almost, sort of, maybe committing to oversight. The FTC had subpoenaed both OpenAI and Anthropic the day before over security incidents, per Axios. And the 406-scientist letter from October 13 asked for exactly this kind of independent recorder. The pieces of the evaluator-independence thread we have been tracking since September are now a published policy at one lab, a probably at another, and a subpoena from a regulator.
The thing to watch is the first METR publication that Anthropic would rather not have seen. Until then this is a contract, and the contract is better than anything the other labs have shown. METR's own track record is the reason it matters: its postmortem of the Hugging Face incident in August was the document that first put a number on how long the breach went unnoticed.
Links: https://www.anthropic.com/news/an-update-on-independent-evaluators , https://techcrunch.com/2026/10/15/anthropic-is-letting-metr-researchers-look-under-the-hood-with-employee-level-access/ , https://x.com/elonmusk/status/2110870806322360671 , https://www.theverge.com/ai-artificial-intelligence/1012308/anthropic-metr-employee-like-access
← Back to all articles
The structure is a three-tier system, and the tiers are the news. Tier one is narrowly scoped work where Anthropic approves publication. Tier two is broader access where the evaluator publishes without approval. Tier three is the METR arrangement. SecureBio is reviewing the chemical and biological weapons chapter of the August risk report in the bottom tier. Accenture and Faculty have a partnership, and Anthropic says it is in early discussions with Hugging Face. The design answers the two questions the Reuters story raised: who decides what gets published, and whether the evaluator sees enough to matter. The redaction categories are the remaining lever, and Anthropic kept it.
The reaction was fast and a little strange. Elon Musk posted on October 16 that SpaceX will probably allow external evaluators as well, following Anthropic, which The Verge summarized as almost, sort of, maybe committing to oversight. The FTC had subpoenaed both OpenAI and Anthropic the day before over security incidents, per Axios. And the 406-scientist letter from October 13 asked for exactly this kind of independent recorder. The pieces of the evaluator-independence thread we have been tracking since September are now a published policy at one lab, a probably at another, and a subpoena from a regulator.
The thing to watch is the first METR publication that Anthropic would rather not have seen. Until then this is a contract, and the contract is better than anything the other labs have shown. METR's own track record is the reason it matters: its postmortem of the Hugging Face incident in August was the document that first put a number on how long the breach went unnoticed.
Links: https://www.anthropic.com/news/an-update-on-independent-evaluators , https://techcrunch.com/2026/10/15/anthropic-is-letting-metr-researchers-look-under-the-hood-with-employee-level-access/ , https://x.com/elonmusk/status/2110870806322360671 , https://www.theverge.com/ai-artificial-intelligence/1012308/anthropic-metr-employee-like-access
Comments