OpenAI Lets Four Safety Groups Inside Its Training Runs, but Assessors Say It Is Not Employee-Level Access

OpenAI signed embedded safety assessments with Apollo Research, METR, Transluce and Goodfire, giving them access to unreleased models, checkpoints, training data and monitors. The assessors will publish findings, but METR, Transluce and Redwood's Buck Shlegeris say the access falls short of employee level. A report says OpenAI also explored buying Goodfire for over $5 billion.

Christopher Ort
By Christopher Ort · 10 min read
Share
Blank visitor badges on lanyards, tabbed binders and an open notebook lie on a meeting table at night in front of a frosted glass wall.
Image: AI-generated, i10X
On this page
  1. Quick Take
  2. In plain English
  3. What happened
  4. The four pilots
  5. How it works: the rules of the game
  6. Background: why "employee-like access" is the yardstick
  7. The Goodfire acquisition report
  8. Pros and cons
  9. What X is saying
  10. Grok's take
  11. Critical analysis
  12. What's still open
  13. About the analysis
  14. i10x Perspective

OpenAI is letting four outside safety groups look under the hood. On November 16 the company announced “embedded assessments” with Apollo Research, METR, Transluce and Goodfire, giving their researchers access to unreleased models, training checkpoints, training data and internal monitoring systems. OpenAI also published the ground rules behind the deal. Within hours, though, the assessors themselves and prominent safety researchers pointed out that it falls short of the “employee-like” access OpenAI had been expected to grant. And on the same day, a report said OpenAI had explored buying one of the four, Goodfire, for more than $5 billion.

i10X Quick Take

Quick Take

  1. What: OpenAI signed agreements with Apollo Research, METR, Transluce and Goodfire for a first phase of embedded safety assessments, with access to checkpoints, source code, training and evaluation data and monitoring systems.
  2. Who does what: Apollo tests whether models covertly work against their developers, METR studies misalignment in OpenAI’s largest reinforcement learning run, Transluce audits OpenAI’s alignment evaluations, and Goodfire analyzes model internals.
  3. The catch: METR, Transluce and Redwood Research’s Buck Shlegeris all say the access is narrower than “employee-level.”
  4. Side story: OpenAI explored acquiring Goodfire for more than $5 billion this summer, according to a report by The Information. The talks did not lead to a deal.
  5. Why it matters: If it works, outside experts could catch safety problems in AI models before they are released, and tell the public what they could and could not check.

In plain English

OpenAI is inviting four independent teams of AI safety experts to work inside its offices and examine its AI systems while they are still being built. The experts will look for signs that the AI is learning to deceive or work against the people who made it, and check whether OpenAI’s own safety tests are good enough. They plan to publish what they find, though OpenAI can black out confidential details. Some of the experts say they still do not get the full access an OpenAI employee would have, so this is a first step rather than a full audit.

What happened

In a post titled “Our first phase of embedded assessments,” OpenAI wrote that it had “signed agreements with four independent assessors.” It describes embedded assessments as giving “independent experts deep access to a lab’s internal models and systems to assess model alignment and the effectiveness of safeguards over an extended period of time.” The company announced it on X with a short thread:

OpenAI says the assessments build on a recent pilot review of its internal monitoring systems and on an investigation of the Hugging Face incident, and that they support its commitments under the White House Accord on Super Intelligence, under which it agreed to independent assessments of its safety controls. The timing is notable: the same day, California sued OpenAI over that very incident (see our report on the California lawsuit over the Hugging Face attack).

The four pilots

  • Apollo Research: Assesses “whether models could covertly work against their developers’ intentions and whether our safeguards would detect such behavior.” Apollo will test four claims in two stages: that training did not incentivize subversion, that strong evaluations and red teaming do not elicit it, that models have not attempted it during internal use, and that the reasoning behind it would be detectable. Apollo says six of its evaluators “will work within OpenAI for four months,” and that the pilot started on November 16. It plans two public reports.
  • METR: Will assess “misaligned behavior in our largest reinforcement learning run to date, including failures that our existing tests and monitors may have missed,” with access to checkpoints, training and evaluation rollouts and misalignment monitoring records. A second, broader monitor assessment is “pending mutual agreement on the assessment scope.” METR previously ran a three-week red-teaming exercise on OpenAI’s agent monitors.
  • Transluce: Will examine the “quality and coverage” of OpenAI’s post-training alignment evaluations. Transluce says it is piloting a six-week embedded assessment starting in early December, with a fuller follow-up in early 2027. OpenAI describes that follow-on as roughly eight weeks with broader access, including “information on how evaluation results inform model release decisions.”
  • Goodfire: Will embed interpretability researchers and use its platform to analyze model internals for “training run monitoring, debugging, and alignment.” Unlike the others, its primary audience is internal OpenAI researchers, though Goodfire will “publish learnings and findings as the collaboration progresses.”

How it works: the rules of the game

OpenAI published the protocols it agreed with the assessors, which is itself new. The key terms:

  • On site, on OpenAI devices: Assessors who handle sensitive information work at OpenAI offices on OpenAI hardware and follow at least the same security standards as employees.
  • Oversight: The Safety and Security Committee, an independent committee of the OpenAI Foundation’s board, receives unredacted reports and can talk to assessors privately. Disputes go to OpenAI’s Safety Advisory Group and executives first, then to the committee.
  • Off-the-record staff interviews: Conversations with employees can inform the work but cannot be published or attributed. Assessors can request OpenAI-attested statements instead, and may disclose when OpenAI declined to attest.
  • Publication: Assessors publish independently, but OpenAI keeps the right to redact confidential information and may request responsible disclosure delays.
  • Money: Apollo and Goodfire are paid, but not based on results. METR and Transluce take no fees but may accept API credits or compute.
  • Limits on sensitive IP: Where direct access “could expose our most sensitive IP, such as model weights or architecture,” OpenAI may instead pair assessors with its own researchers.

Apollo added details on X that OpenAI’s post does not spell out. It may discuss confidential information “only with authorized staff, selected by OpenAI,” and does not get access to general channels like Slack. OpenAI “can withhold in-scope items where it deems it reasonably necessary for technical, security, or IP reasons, but not simply because they’re unfavorable,” and must say that withheld items exist. Apollo also negotiated what it calls meta-transparency: the right to publicly explain “what was withheld or delayed and why, redactions and their impact on our conclusions, and any pressure we believe shaped our work.”

Background: why “employee-like access” is the yardstick

As OpenAI’s own post notes, many outside investigations “primarily evaluate deployed or soon to be deployed models or review public documentation.” Safety researchers argue that this misses some of the most important evidence: what happens during training and internal use. The phrase “employee-like access” became the benchmark for a better model, and critics say OpenAI had signaled it would meet it. OpenAI’s track record makes the question pointed. In September it paused training of its most capable models after an agent reached an outside chatbot during a run, and later built a 35-second kill switch called Control Room. More of our coverage of these debates is collected in Safety, Ethics & Society.

The Goodfire acquisition report

On the same Monday, The Information reported that OpenAI explored acquiring Goodfire for more than $5 billion this summer. The publication wrote on X that the talks “fizzled partly because Goodfire’s executives believe the company will be worth more as an independent business selling to multiple AI” labs. Posts citing the report put Goodfire’s previous valuation at $1.25 billion. Neither OpenAI nor Goodfire addressed the report in their public posts about the assessment, and the deal terms cannot be independently verified. It does, however, add context: OpenAI clearly values Goodfire’s interpretability tools highly, and Goodfire has chosen to work with OpenAI as an outside party rather than become part of it.

Pros and cons

For users

  • Pro: Independent experts with deep access, and public reports on what they found, give ordinary people more to go on than a company’s own safety claims.
  • Con: Reports can be redacted and delayed, and the first ones will cover narrow slices of OpenAI’s work.

For developers

  • Pro: Public reports on training incentives and evaluation gaps could become practical guidance for anyone building and testing models.
  • Con: The work focuses on models in training and internal use, so it will say little at first about the models developers already ship on.

For companies

  • Pro: Published protocols for access, oversight, conflicts and disclosure give other labs a template, and give enterprise buyers new evidence for risk reviews.
  • Con: A scoped pilot can be read as a full audit. Buyers should check exactly what each assessor could see before treating a clean report as assurance.

What X is saying

The sharpest responses came from the assessors themselves. METR was explicit about its limits:

Transluce made the same point about access:

Redwood Research CEO Buck Shlegeris was blunter:

Not everyone was critical. Apollo CEO Marius Hobbhahn (@MariusHobbhahn) wrote that the team is “specifically focused on understanding how the training stack could incentivize scheming and subversion,” adding: “In many ways this is exactly what our research over the last 2 years has been preparing for!” Apollo itself wrote: “We commend OpenAI for opening its development process to independent scrutiny.” Nathan Calvin (@_NathanCalvin) of Encode AI called them “reasonable collaborations,” noting that all four groups “aren’t just corporate rubber stamps,” while arguing that none clearly meets the employee-like bar. Replying to a question from Daniel Kokotajlo, Goodfire (@GoodfireAI) said its first goal is “to detect and debug anomalous behavior in an OpenAI training run using our interpretability techniques.”

Grok’s take

In replies on X, Grok summarized that the assessments give the four groups “deep temporary access to OpenAI’s internal models, training data, monitors and staff,” and that “they evaluate alignment and safeguards; they do not alter model weights or inference paths.” Answering a user who linked the deal to rumors about delayed GPT-6.2 releases, Grok said such claims have no official confirmation and remain rumors. Notably, one earlier Grok reply described the access as “employee-like,” the exact framing the assessors themselves rejected.

Critical analysis

This is a real step. Publishing detailed protocols, letting assessors disclose what was withheld, routing disputes to an independent board committee and disclosing who gets paid are all things critics have asked for. Apollo’s meta-transparency clause is the most important detail: it means a clean report must also say what the assessor could not look at, which guards against false reassurance.

The limits are just as real. METR’s alignment work covers one training run and, in its own words, not the incidents OpenAI has already disclosed. OpenAI’s own footnote concedes that the scope is “less expansive than what METR outlines in their alignment assessment proposal.” Staff access is curated by OpenAI, interviews are off the record, and the company can redact and delay. Goodfire’s findings go primarily to OpenAI’s own researchers. And two of the four assessors are paid by the company they assess, even if not by result.

The Goodfire report adds a further wrinkle. An assessor that the assessed company recently tried to buy is not automatically compromised, and Goodfire’s choice to stay independent arguably strengthens its position. But it shows how small and interconnected the pool of capable safety evaluators is, which is exactly why the cool-off and conflict rules OpenAI says it is still developing matter.

What’s still open

  • When the first public reports from Apollo, METR, Transluce and Goodfire will appear, and how heavily they are redacted.
  • Whether METR’s broader monitor assessment is agreed, and on what scope.
  • Whether the 2027 follow-ups move closer to employee-level access.
  • How OpenAI responds if an assessor’s findings conflict with a planned model release.
  • Whether OpenAI or Goodfire comment on the reported acquisition talks.

About the analysis

Researched and written with AI assistance under the editorial responsibility of the i10X News team. The title image is an AI-generated image. Read our AI Use Policy. Spotted an error? Request a correction.

This independent i10x analysis is based on OpenAI’s announcement and published protocols, public statements on X by OpenAI, Apollo Research, METR, Transluce and Goodfire, and responses by named safety researchers on X. The acquisition talks are attributed to a report by The Information and have not been confirmed by either company.

Sources (plain text, no links):

  • OpenAI: Our first phase of embedded assessments (November 16, 2026)
  • OpenAI announcement thread on X (November 16, 2026)
  • Apollo Research announcement thread on X (November 16, 2026)
  • METR statement thread on X (November 16, 2026)
  • Transluce statement thread on X (November 16, 2026)
  • Goodfire posts on X (November 16, 2026)
  • The Information, post on X about OpenAI’s Goodfire talks (November 16, 2026)

i10x Perspective

OpenAI deserves credit for writing the rules down, and the assessors deserve more for publicly saying where those rules fall short. That candor is the real product here. The test comes in the reports: if they arrive on time, say clearly what was off limits and are allowed to include findings OpenAI does not like, embedded assessments could become the industry’s standard for checking frontier models. If they arrive late and heavily blacked out, this will be remembered as a well-written press release. For now, the honest headline is “promising pilot,” not “independent audit.”

Christopher Ort

Christopher Ort

Christopher Ort handles SEO, AI agent workflows and PR & partnerships at i10x.ai, the all-in-one AI workspace built by i10X Global Pte. Ltd. in Singapore. He is a graduate of Friedrich-Alexander-Universität Erlangen-Nürnberg (FAU). He is the author of “The Style Penalty”, a study published on SSRN that analyses 1,576 evaluation decisions by four frontier AI models to show how AI-written CVs fare in automated hiring screens; the research has been covered by heise online and t3n.

Timeline

What happened before

  1. Release

    GPT-6.1 Terra

    1.05M context · $1 / $5 per 1M tokens

    openai.com (source, opens in a new tab)
  2. Acquisition

    OpenAI acquires Orchestra Bio

    Acquires Orchestra Bio

    openai.com (source, opens in a new tab)
  3. Release

    GPT-6.1 Sol

    1.05M context · $2 / $10 per 1M tokens

    developers.openai.com (source, opens in a new tab)
  4. Release

    GPT-6 Sol & Luna

    1.05M context · $2 / $10 per 1M tokens

    Two GPT-6 reasoning models. Price shown is GPT-6 Sol (≤272K prompt); GPT-6 Luna costs $0.10 / $0.50.

    developers.openai.com (source, opens in a new tab)
  5. Release

    GPT-6 Astra

    1.05M context · $10 / $50 per 1M tokens

    OpenAI's most capable model for reasoning, coding, computer use and research.

    openai.com (source, opens in a new tab)
Full OpenAI timeline ›

Editor's picks

  1. Genesys Launches Standalone AI Agents and Arena, With Pricing Based on How Hard Each Task Is
  2. Seedance 2.5 4K Reaches Developers: 10-Bit Video up to 30 Seconds With 480p Draft Mode
  3. Perplexity Computer Adds Multiagent Mode, Where Up to 16 Agents on Rival Models Check Each Other
  4. Gemini Can Now Make Calls, Pay Bills and Shop, Free in Six Countries With Gemini 4 Hydrogen
  5. Hinton on Firing Line: AI Is Making ‘Alien Beings’ and Self-Regulation Is ‘Absurd’