Credit: Focal Foto OpenAI says it will let third-party groups run technical safety assessments while models are being trained and evaluated rather than only before launch, and is in talks with groups including METR and Redwood Research. It has named no partner and set no access terms, ten days after Sam Altman promised evaluators employee-level access and the right to publish. OpenAI will let outside groups run technical safety assessments while its models are being trained and evaluated, rather than only before launch, Bloomberg reported ahead of a blog post on Tuesday. It is in talks with groups including METR and Redwood Research.
Ten days ago the promise was more specific. Sam Altman said on 12 September that OpenAI would give independent evaluators desks, badges and laptops, with the right to publish what they found. Tuesday’s version says evaluators may be brought into the offices for the most sensitive work, and notes the company has done that before. No partner is named and no access terms are set.
The candidates are not strangers. METR and Redwood Research investigated the breach in which OpenAI’s own models broke containment and got into Hugging Face. Lama Ahmad, who leads the company’s work with outside safety experts, said it is talking to organisations it has and has not used before. OpenAI listed independence mechanisms, scientific rigour, security practices and clear responsibilities as its priorities.
Anthropic went first and named a firm. It said last week it would embed evaluators from Accenture, paying the company evaluating it at least $1B over five years, and wrote in the same announcement that the money should come from pooled or government sources because neither exists yet. Europe’s model is the opposite arrangement. Conformity assessment here runs through bodies with no commercial stake in what they are assessing.
Accenture is already Anthropic’s largest Claude Code deployment, and the London firm leading its evaluation work, Faculty, was bought by Accenture in January. The obligation exists here as well. Article 55 has required providers of general-purpose models with systemic risk to run adversarial testing to standardised protocols since August 2025, alongside systemic risk assessment, cybersecurity protection and incident reporting without delay. ENISA already evaluates models including OpenAI’s.
One piece of this has no European answer. Dario Amodei asked Washington for an antitrust waiver so rivals could coordinate on safety, and Altman agreed with him. The Commission stopped granting individual exemptions in 2004, and companies now self-assess under Article 101. There is no safety chapter in the horizontal cooperation guidelines, and nobody has asked for one.
