Anthropic, OpenAI pledge to embed outside safety evaluators

On Sept. 12, 2026, Anthropic CEO Dario Amodei published “We Must Pace the Frontier” and unilaterally committed Anthropic to embedded third-party evaluators — ongoing, employee-like access so outsiders can check safety claims against training pipelines and processes, not only finished weights. OpenAI CEO Sam Altman said the same day that OpenAI would match the access standard.

This is lab self-rule, not Congress. No statute requires the desks. The commit is the verifiability layer Amodei says any pacing deal needs.

The rule in English

Step 1 — Embedded Evaluators (Amodei essay). Each frontier AI company gives ongoing, employee-like access to a team of third-party evaluators (essay names METR as an example). Role: verify adherence to safety practices and commitments; report incidents; help assess alignment of training pipelines and processes, not just completed models. Banking “supervisors” embedded with employees is the stated precedent.

Anthropic’s near-term invite list (essay): desks in offices, access badges, company laptops; workspaces, tools, and permissions mostly comparable to internal risk-assessment teams (exceptions for law, contracts, customer/partner private data); a contract with the right to publish key findings on risk levels, incidents, practices, and access received — without Anthropic editorial control, subject only to narrow redactions (security-sensitive, privileged, commercially sensitive, or third-party confidential). Reviewers may say publicly if a redaction removed something important to their conclusions.

OpenAI match (Altman, Sept. 12, via contemporaneous reports of his X post): “Committing to having independent evaluators with employee-like access is a great idea, and we will do the same. We’ll have more to share soon.” OpenAI has not yet published desks-badges-publish terms matching Anthropic’s essay detail.

What this is not. Steps 2–3 in the essay (democratic industry coordination; global coordination with authoritarian states) are asks, not commits. Amodei’s call to slow capability gains remains voluntary unless governments require it. Trump’s same-week All-In / Truth Social posture rejects the slowdown politics.

Who wrote it / who it binds

Writers: Amodei (Anthropic unilateral commit in the essay); Altman (OpenAI public match on X, per secondary coverage). Binds Anthropic’s stated intent and OpenAI’s stated access principle once operationalized. Does not bind Google DeepMind, xAI, Meta, Microsoft AI, or startups. Does not create federal auditor status for METR or any named shop.

Who it exempts

Competitors who do not match. Narrow Anthropic exceptions for law, contracts, and third-party confidential data. Altman’s “more to share soon” leaves OpenAI’s redaction, publication, and pipeline-depth rules undefined in public as of publication.

What a founder, parent, or worker should do differently

Founders at frontier labs: treat employee-level pipeline access + publish-without-editorial-control as the bar Anthropic set; matching Altman on a slogan is cheaper than matching Anthropic’s contract terms.

Enterprise buyers: ask vendors whether third-party evaluators have training-pipeline access or only pre-release model snapshots.

Parents and workers: this is lab governance theater until evaluators are named, seated, and publishing. It is not a school or platform safety statute.

What’s next

• Named evaluator org(s), start date, and first public finding (or first public complaint about redaction).
• Whether OpenAI’s “more to share” matches Anthropic’s publish-without-control clause.
• Whether Congress or agencies convert embedded access into a legal duty — Amodei asks them to; the White House has not.
• Whether other labs match access depth, not just the phrase.

Source: Amodei, “We Must Pace the Frontier”.