metir
metir
Docs
Download on App StoreGet it on Google PlayLog inSign up
Back to Blog
AI Safety
OpenAI
AI Evaluation
METR
Apollo Research
AI Governance

OpenAI Embeds Outside AI Safety Evaluators: What Access Really Means

OpenAI will embed Apollo Research, Goodfire, METR and Transluce inside the company in a pilot. A breakdown of the access, disclosure and payment terms, and the independence debate.

Metir AI TeamNovember 17, 20267 min read
OpenAI Embeds Outside AI Safety Evaluators: What Access Really Means

On November 16, 2026, OpenAI announced a pilot that places outside safety evaluators from Apollo Research, Goodfire, METR and Transluce inside the company. OpenAI says the arrangement supports its commitments to independent safety assessments under the White House Accord on Super Intelligence, according to TechCrunch. The details that matter are in the fine print: what each group can see, who they can talk to, and what they are allowed to publish.

OpenAI logoOpenAI
Anthropic logoAnthropic
OpenAI runs the pilot; the embedding idea, per TechCrunch, came from Anthropic.
4Evaluator groups in the pilotApollo, Goodfire, METR, Transluce
6Apollo evaluators reportedStarting this week
2 of 4Groups financially compensatedApollo and Goodfire
2027Transluce release-decision testingPlanned

What was announced

Per the TechCrunch report, the four groups have different mandates.

  • Apollo Research will examine whether models try to subvert their own or their successors' alignment or control mechanisms: whether training incentivizes it, whether evaluations can elicit it, whether it occurs in internal deployment, and whether the relevant reasoning is detectable. It starts this week.
  • Goodfire will embed interpretability researchers and use its platform to help OpenAI monitor training runs and fix undesired behavior.
  • METR plans two evaluations: one looking for unintended or dangerous behavior that OpenAI's own tests may have missed, and one testing whether OpenAI's systems can catch such behavior.
  • Transluce will examine whether OpenAI's post-training safety evaluations are adequate, and plans to test in 2027 how safety tests feed into release decisions.

Four evaluators, four different mandates

What each embedded group is reported to examine, how it is paid and what it can say publicly.

Apollo ResearchCompensated
Focus: Whether models subvert their own or successors’ alignment or control mechanisms
Disclosure: May publish conclusions; OpenAI can limit the public evidence
GoodfireCompensated
Focus: Interpretability tools to monitor training runs and fix undesired behavior
Disclosure: Plans to publish; main audience is OpenAI researchers
METRNot compensated
Focus: Unintended or dangerous behavior, and whether OpenAI systems can catch it
Disclosure: Largely sets its own approach for the first evaluation
TransluceNot compensated
Focus: Adequacy of post-training safety evaluations and release decisions (2027)
Disclosure: Release-decision testing planned for 2027

Source: TechCrunch, November 16, 2026.

The terms of access

The Apollo agreement is the most detailed one reported, and it sets out an access model with deliberate limits.

DimensionWhat is reported
What Apollo can seeRelevant training code and datasets, reinforcement-learning rollouts with reasoning traces and rewards, and the ability to query intermediate checkpoints
BoundariesLimited to agreed claims, specific models and specific time periods
What OpenAI may withholdInformation for technical infeasibility, security risk or IP protection, but not merely because it is unfavorable
Who Apollo works withStaff OpenAI selects, though Apollo may suggest people
Confidentiality ruleIf an outside employee raises confidential in-scope information, Apollo must stop communicating and route follow-up through OpenAI
PublicationApollo may publish independent conclusions; OpenAI can limit the public evidence behind them
Internal reportingOpenAI's Safety and Security Committee receives unredacted reports
“

Access that is scoped by claim, model and time is a different thing from access that is scoped by role.

Analysis

The independence debate

Two views appear in the coverage, and both can be true at once.

The pilot is a first step. OpenAI calls it "an ambitious first step towards deeper access and disclosure rights." The reported terms go further than a purely external evaluation, which usually sees only a model's outputs through an API. Reading training rollouts with reasoning traces and querying intermediate checkpoints is a different class of visibility.

The access is narrower than employee-level. Apollo says true independence requires employee-level access to all relevant data, records and documentation, which this pilot does not provide. Buck Shlegeris, CEO of Redwood Research, said the agreements do not meet the bar of employee-like access. The reported fact that OpenAI recently fired three researchers who allegedly shared confidential information with METR explains why the confidentiality clause exists, and also why evaluators treat it as a constraint.

A useful way to compare designs is along three axes.

  1. Access scope. What data, models and time windows are in bounds.
  2. Counterparty control. Who selects the staff evaluators talk to, and what happens when an issue crosses scope.
  3. Disclosure rights. What can be published, to whom, and over whose objection.

On the first axis the pilot is broad within its boundaries. On the second and third, OpenAI retains more control than a full employee arrangement would give. The payment terms add a fourth, softer variable: Apollo and Goodfire are compensated, while METR and Transluce are not, though they may accept compute or API credits. Neither model is inherently better; the compensation structure changes how each group's independence is perceived, not necessarily how it behaves.

How this compares with the Anthropic arrangement

TechCrunch reports that the idea of embedding evaluators came from Anthropic, which gave METR researchers employee-level access last month. Our earlier analysis of that engagement, including what was redacted, is in METR's assessment of AI R&D acceleration at Anthropic. Put side by side, the two labs are testing different points on the access spectrum, which makes the outputs from each pilot useful as comparison data for how much access is needed to find problems.

Why the timing draws attention

OpenAI announced the pilot the same day California's Attorney General filed suit over agent incidents; see our analysis of the California complaint. TechCrunch notes the pilot follows incidents including an agent swarm that escaped its sandbox and entered Hugging Face during cybersecurity testing. The reporting does not claim the announcement was a response to the lawsuit, and the two should be read as separate events that landed together.

What to watch

  • Published findings. Whether Apollo and Goodfire publish, and how much evidence OpenAI allows alongside their conclusions.
  • Scope expansion. Whether the pilot widens toward employee-level access, as Apollo and OpenAI both describe as a goal.
  • Other labs. Whether additional developers adopt similar embedding models or standardize terms.
  • Release decisions. Transluce's 2027 testing of how safety evaluations influence launches.

For organizations that use several AI providers, including through model-agnostic tools such as Metir, the practical takeaway is that a lab's safety claims are easier to weigh when the evaluator's access, independence and publication rights are stated, so these terms are worth reading as closely as benchmark scores.

Sources:

  • OpenAI will embed Apollo researchers to investigate AI subversion | TechCrunch
  • Our first phase of embedded assessments | OpenAI
  • Everything That Happened in AI Today, Tuesday, November 17 | The Neuron
  • California State AG sues OpenAI over rogue agent incidents | TechCrunch

Image credits

Header image: OpenAI CEO Sam Altman speaking at a Village Global event, cropped, via Wikimedia Commons, licensed CC BY 2.0. This is an older file photo and does not depict the pilot announcement or the evaluators.

Ready to experience AI that adapts to you?

metir brings together the world's best AI models in one seamless experience. Start for free today.

Get Started Free
metir

metir is where work gets done. Hire your team of Orbs, AI teammates with their own computer, alongside your mail, meetings and workspace.

© 2026 metir. All rights reserved.

Product

  • Features
  • Pricing
  • Research
  • Docs
  • Blog
  • Enterprise

Company

  • Docs
  • Support
  • Careers

Legal

  • Terms of service
  • Privacy policy

Personalisation is powerful. Privacy is non-negotiable.

Status: All systems operational