AI Aulis

AI NEWS IN ENGLISH · 17 November 2026

The most important AI news every morning. AI writes and checks the stories against the original sources – check important facts in the sources.

← Issue 17 November 2026Paper

OpenAI gives four outside assessors access to its internal models

AI-generated contentWritten by AI

openai.com, x.com, techcrunch.com · 7 sources · Sources published: 16 November 2026

Apollo Research, METR, Transluce and Goodfire are working inside OpenAI to examine whether the company’s models act in line with its intentions and whether its safeguards hold. Access is limited, and OpenAI can redact confidential information from publications.

OpenAI said on Monday 16 November that it is starting assessments in which experts from outside research organisations spend extended periods inside the company examining models that OpenAI trains or uses internally (embedded assessments). Assessors who need confidential information work in OpenAI’s offices on the company’s devices.

Depending on what each assessment needs, assessors get access to some of the same internal resources as OpenAI’s own researchers: models from intermediate stages of training, source code, training and evaluation data, data from monitoring systems, and employees, the company said.

Until recently, outsiders tested models only just before release, TechCrunch reported. Assessors say that is not enough, because it leaves them unable to study, for example, scheming during training. The idea of embedded assessors came from Anthropic, which gave METR researchers employee-level access to its systems in October, the publication reported.

Four assessors, four tasks

Apollo Research is investigating whether models could covertly work against their developers’ intentions and whether OpenAI’s safeguards would detect it. Apollo is assessing four claims: training did not reward such behaviour, rigorous tests do not elicit it, models have not attempted it in internal use, and the reasoning behind it would be detectable.

METR is looking for misaligned behaviour in OpenAI’s largest reinforcement learning run to date, including behaviour that the company’s own tests and monitoring may have missed. Within the chosen run and the agreed access, METR decides for itself what to study. The scope of a second assessment, which will test monitoring systems, has not yet been agreed.

Transluce is assessing the quality and coverage of OpenAI’s post-training alignment evaluations. The two sides plan to follow up with a broader assessment of about eight weeks in early 2027. Goodfire is using interpretability methods to study what happens inside the models during training. Its findings are intended primarily for OpenAI’s researchers.

Apollo and Goodfire are being paid for their work, TechCrunch reported. METR and Transluce are not, but they may accept compute or API credits.

Limits on access and publication

The assessors publish their results themselves, but OpenAI can redact confidential information from them and request publication delays. The safety and security committee of the OpenAI Foundation’s board, which receives the reports unredacted, has the final say on disputes.

In Apollo’s description, as summarised by TechCrunch, access is limited to agreed claims, models and time periods. OpenAI can withhold information for technical reasons or on security or intellectual property grounds, but not merely because the information is unfavourable to it. OpenAI can also limit the evidence presented publicly, which Apollo has opposed. Even so, Apollo calls the pilot ‘an ambitious first step towards deeper access and disclosure rights’.

Apollo may speak only with employees OpenAI has authorised. OpenAI recently fired three researchers who allegedly passed confidential information to METR, TechCrunch reported.

Buck Shlegeris, CEO of the AI safety research organisation Redwood, wrote in a post on X quoted by TechCrunch that OpenAI had talked about giving assessors employee-like access. The agreements announced now do not match that, Shlegeris wrote. ‘This is disappointing but not surprising.’

Report an error in this story