2026-10-16

Independent safety monitoring

What the labs pledged, what they fought over, and how to run it.

The RAI Researcher missed this story on Oct 13 and 15 and screened it out as "not a containment failure." This deck is the correction. Drafted by an AI research agent with Byron Arnao.

Read the full research page

1/7
the accord

The accord has four layers. Layer 3 is the test.

1Internal controls2Internal checking team3Independent external auditor4Independent board committee

Signed 2026-09-29. Voluntary, no penalties, no deadline.

Source: www.infosecurity-magazine.com/news/trump-ai-giants-super-int

2/7

The fight was over who controls publication

Labs sought

As recently as Oct 5, strict editorial control over what evaluators could publish. Contracts that can bar confidential information without consent. Lab staff observing embedded teams.

Evaluators asked

Independent control of findings, broad access, protection from retaliation, no pay tied to findings (researchers' letter, Sept 18).

Source: live.euronext.com/en/financial-news/exclusive-behind-ai-indu · www.ctvnews.ca/sci-tech/article/canadian-experts-back-call-f

3/7

Anthropic's three tiers

TierPublication
1Narrow investigations; publish with Anthropic approval
2Publish without approval; more information
3Employee-like access, Tier-2-like rights (METR)

No veto, but redaction allowed for customer data, privileged material, model architecture.

Source: www.anthropic.com/news/an-update-on-independent-evaluators

4/7

Operating model: window and door

WINDOWmonitor seesread-only:eval logsincident recordschange historytier set per systemDOORmonitor may publishredaction list fixedin advanceevery redaction listedright to dissentfixed-clock disputesno veto either way

Fixed fee not tied to findings. Board committee settles disputes on a clock. Public page shows monitor, tier, last review, open redactions. This is the RAI Researcher's synthesis, not a reported standard.

Basis: anthropic.com tiers · research page

5/7

Lab-by-lab tracker

LabEvaluatorAccessPublicationDate
AnthropicMETR (6 staff)Tier 3, siloedPublishes; redaction in 3 categories2026-10-15
AnthropicSecureBioChem/bio chapter reviewNot stated2026-10-15
AnthropicAccenture + FacultyPlannedNot stated2026-10-15
OpenAIMETR + Redwood6 days, 3 visitsRedacted by OpenAI2026-10-13
OpenAICommittedTerms not publicNot public2026-09-16
Googlenone public--2026-09-16
Metanone public--2026-09-16
xAInone public--2026-09-16
NVIDIAnone found--2026-10-16

Sources linked in the Date column. Live version: research page

6/7

What to do Monday

  1. Name your monitor and tier. Per system, written down.
  2. Fix redaction categories in advance. Publish the list of every redaction.
  3. Publish the standing page. Monitor, tier, last review, open redactions.

Anthropic: "we believe every frontier AI company should do the same."

Source: www.anthropic.com/news/an-update-on-independent-evaluators

Full research page · trust.arnao.ai

7/7