TellWorldAI World Archive
Ask中文

Home›News›Partnership

PartnershipGenerative AI for everyone

OpenAI begins embedded safety assessments with four outside evaluators

On 16 November 2026 OpenAI said its first phase of embedded safety assessments involves Apollo Research, METR, Transluce and Goodfire. The evaluators will work inside OpenAI to test model alignment, safeguards and monitoring.

Four small inspection drones docked around a glass sphere, scanning the glowing neural lattice inside.

Background

Safety assessments check whether a model behaves as its developer intends and whether the safeguards around it work. Independent evaluators, meaning outside groups not employed by the developer, are one way to get those checks done. OpenAI's approach here puts such evaluators inside the company for a period.

What happened

On 16 November 2026 OpenAI said that its first phase of embedded assessments involves Apollo Research, METR, Transluce and Goodfire.[1] OpenAI describes the assessments as independent evaluations of model alignment, safeguards and monitoring systems.[1] Alignment is the degree to which a model's behavior matches the goals its developers set for it. Monitoring systems watch what a model does so that problems can be noticed.

Key details

TechCrunch reports that the evaluators will be embedded inside OpenAI, and that the Apollo pilot begins the same week.[2]

Apollo's pilot will examine whether models try to subvert their own or their successors' alignment or control mechanisms.[2] Subverting such mechanisms would mean a model working against the human oversight meant to keep it on track, which is why this question is a central one for safety researchers.

According to TechCrunch, Apollo cannot follow up on confidential information raised by OpenAI employees outside its approved group. Apollo also says it needs OpenAI's agreement to publicly explain the conditions and limits of its work.[2]

OpenAI says the work supports its commitments to rely on independent assessments of safety controls under the White House Accord on Super Intelligence.[2]

Impact

The arrangement shows what embedded evaluation looks like in practice: outside researchers gain access to test a lab's systems from the inside, and the terms of that access, including what can be disclosed, shape what the public learns. Readers can compare the company's description of the assessments with the limits that the evaluator itself describes.

Sources

  1. 1
  2. 2

Cite this page: OpenAI begins embedded safety assessments with four outside evaluators. TellWorld, https://www.tellworld.net/news/openai-embedded-assessments-2026. Updated 2026-11-18

Related news

View all