Cybersecurity & Privacy

OpenAI Opens Internal AI Safety Tests to Four Outside Assessors

OpenAI is giving four outside assessors deeper access to internal models and systems, with Apollo planning a six-person, four-month pilot.

By Marcus Lee Edited by Samantha Reed Published: Updated:
OpenAI Opens Internal AI Safety Tests to Four Outside Assessors
OpenAI is giving four outside assessors deeper access to internal models and systems, with Apollo planning a six-person, four-month pilot. Photo: Zulfugar Karimov / Unsplash

Key Notes

  • OpenAI has signed agreements with four outside organizations to assess internal model alignment and safeguards.
  • Apollo plans a six-person pilot lasting four months, with public reports that explain access and disclosure restrictions.
  • The program has not yet established safety findings, and readers will need to examine each assessment’s scope and limits.

OpenAI has signed agreements with four independent organizations to assess the safety of models during training and internal use. The company announced the first phase on November 16, naming Apollo Research, METR, Goodfire and Transluce.

The assessors will receive access tailored to their assignments, potentially including unreleased model checkpoints, code, training records and monitoring systems. The question is whether outside researchers can test the evidence behind a lab’s safety claims, including behavior that never appears in a public chatbot.

Apollo Sends Six Evaluators for Four Months

Apollo Research separately described a six-person, four-month pilot examining training, evaluations and internal deployment. It will investigate whether models try to undermine alignment or control mechanisms, and whether relevant reasoning would be detectable.

Apollo plans two public reports. Its agreement allows it to explain restrictions on its work and how withheld evidence affects its conclusions. OpenAI may redact sensitive evidence, but Apollo says the company cannot remove its verdicts or request redactions simply because a finding is unfavorable.

The access is scoped, however. Apollo says some arrangements remain to be agreed, expansion is not guaranteed, and security, intellectual-property or technical constraints can limit what it receives. These are terms for an assessment, not a finding that the models are safe.

OpenAI says METR’s initial alignment work will focus on its largest reinforcement-learning run. Goodfire will examine model internals, while Transluce will assess post-training alignment evaluations. The company discloses that Apollo and Goodfire are paid, whereas METR and Transluce do not accept assessment fees but may receive compute or API credits.

Earlier Monitoring Tests Show Why Access Matters

METR’s October report on a separate exercise offers context. Researchers spent three weeks examining OpenAI’s internal monitoring system as it existed on or before July 21. They received code access, monitoring configuration and interviews with staff, producing a detailed internal report.

They identified weaknesses that could let agents avoid or interfere with monitoring. METR stressed that the exercise examined a historical system. OpenAI subsequently said most vulnerabilities had been patched, but METR had not evaluated those remediation claims. The findings should not be presented as a test of the current system.

METR’s broader alignment-assessment proposal asks evaluators to examine both observed incidents and reasons the records might underestimate risk. Missing logs, limited access and a model’s awareness that it is being tested can all affect what an assessment establishes.

The Reports Will Need to Explain Their Limits

Independent scrutiny is most useful when readers can distinguish what researchers checked from what they could not inspect. A public conclusion carries a different weight if its supporting evidence is restricted, its scope excludes important systems or key access arrangements remain unfinished.

AIstify previously reported on a similar move to give outside evaluators deeper access at another frontier lab. OpenAI’s announcement adds a concrete set of organizations and assessment plans. The substantive test will be the findings, the access behind them and the limitations disclosed alongside them.

Disclaimer: AIstify is an independent media brand owned and operated by NuvexMedia LLC, publishing news, research, and insights on artificial intelligence, emerging technologies, automation, and related industries. NuvexMedia LLC invests in and collaborates with companies across the AI, technology, software, and digital innovation sectors. These relationships do not influence AIstify’s editorial coverage, and the publication maintains full editorial independence to provide accurate, timely, and objective information. © 2026 NuvexMedia LLC. All rights reserved. This content is for informational purposes only and should not be considered legal, tax, investment, financial, or other professional advice.

AI & Machine Learning, Cybersecurity & Privacy, News, Regulation & Policy