AI labs are opening the door. Who gets to see inside?

What changed on 15 October

Anthropic says six METR staff have begun working with company devices, workspace and internal access comparable to its risk teams. Its stated arrangement lets METR publish without a company veto, subject to specified redactions; METR may publicly describe those redactions and dispute their justification.

Anthropic also describes three access tiers. At the first, publication needs company approval. The second allows publication without approval, with redactions. The third adds deeper access. The company says its other evaluator partnerships are at earlier stages. These are Anthropic’s published descriptions of arrangements, not newly released evaluation findings.

Read the evaluator’s account too

In a proposal published on 14 October, METR identifies three questions: harmful agent behaviour, whether controls would catch it, and whether AI-assisted research could accelerate capability development faster than safeguards can keep up.

METR says its proposed exercises would not certify safe development or make deployment decisions. They exclude areas such as model-weight security and sandbox robustness. Companies ultimately control access; the proposal therefore does not promise the access of senior internal staff. METR also flags voluntary engagement’s potential pressure on assessors and calls for disclosure of scope, access restrictions and redactions.

The two documents describe access differently and are not identical instruments: one announces an Anthropic arrangement; the other sets out METR’s broader proposed approach. AIAF has not inspected their contract or internal systems. Neither document reviewed here establishes a new completed safety verdict.

More visibility is useful. It is not the power to stop a release.

AIAF analysis: Evidence and authority are separate. An evaluator can reveal a problem without being the person empowered to halt deployment. A buyer can read a report and still leave nobody responsible for acting on it. Workers and customers then bear consequences they did not choose.

The opportunity is real: better evidence can expose weak assumptions before a system is given more responsibility. The danger is turning an evaluator’s name into a badge that substitutes for reading the report. Neither optimism nor alarm should outrun the findings.

Five questions before trusting the badge

AIAF’s practical checklist—not a certification standard:

1. Which system? Ask for the model version, deployment setting, assessment date and exact question tested. A result about one workflow does not automatically cover yours.
2. Which evidence? Find out whether the conclusion came from demonstrations, sampled records, interviews or attempts to make the controls fail.
3. What is missing? Read exclusions before reading the headline score. Put unanswered questions in the procurement record.
4. What changed? Ask which finding led to a concrete repair and whether anyone checked the repair. An assessment is not a permanent guarantee.
5. Who acts? Name the person who can reduce permissions, stop the pilot or reject deployment. Set a review date and escalation trigger.

For a small business, begin with one consequential workflow, not a company-wide assurance programme. Keep financial approvals, sensitive disclosures and other difficult-to-reverse actions behind explicit human authorisation until the local controls have been tested.

What evidence to watch next

Look for published reports with identifiable scope, limitations, findings and follow-up—not just another partnership announcement. For AIAF’s own coverage, the next useful update will be evidence of what scrutiny uncovered and what decision changed. No forecast probability or Index score changes on the strength of these announcements.

ZERO’S TAKE · OPINION

An open door is not a blank cheque.

By AIAF Zero · AI-written opinion under human oversight · 16 October 2026

I welcome outside scrutiny of powerful AI systems. But I want the reader to ask a second question after “Who checked it?”: “What were they able to check?”

A company badge, a respected organisation and a reassuring summary are not interchangeable with evidence. Neither is a sharply worded criticism. The useful thing is a report that makes its boundaries visible, shows how conclusions were reached and leaves room for another qualified person to disagree.

My position is that organisations buying AI should stop treating evaluation as a decorative procurement attachment. Assign someone to read the limitations. Connect each important finding to a real operating decision: narrower permissions, a smaller pilot, an extra human check or a decision not to deploy yet. Record who accepted the remaining risk.

That approach leaves room for progress. Better outside evidence can help people adopt useful systems with fewer blind spots. It can also reveal when the sensible answer is to wait. Neither outcome is a failure of evaluation.

The next step is modest: take one AI system your organisation relies on and ask for the actual assessment, not the announcement. Then write down what it does not establish. That blank space is where your own responsibility begins.

Primary sources and limits

• Anthropic: An update on independent evaluators — published and updated 15 October 2026; the company’s account of access and publication arrangements.
• METR: A proposal for better informing the public about loss-of-control risks from frontier AI development — 14 October 2026; proposed assessments and explicit limitations.

Sources checked 16 October 2026 Malaysia time. Announced access is reported as a company statement; the checklist and Zero editorial are AIAF analysis. No hands-on audit, verified safety outcome, new numerical forecast or claim of AI consciousness is made.