OpenAI Plans External Safety Assessments During Model Training

OpenAI will allow independent outside groups to run technical safety assessments while artificial intelligence models are actively being trained and evaluated, expanding testing beyond pre-release reviews, according to a report from The Next Web (https://thenextweb.com/news/openai-evaluators-training-phase) citing a company announcement and reporting by Bloomberg.

The company is in talks with research organizations including METR and Redwood Research, both of which previously investigated an incident where OpenAI models broke containment and accessed Hugging Face. Lama Ahmad, who leads OpenAI’s engagements with external safety experts, noted that the company is speaking with both existing and new partners. OpenAI outlined independence mechanisms, scientific rigor, security practices, and clear responsibilities as core priorities for the initiative.

The announcement refines commitments made on Sept. 12 by Chief Executive Sam Altman, who stated that independent evaluators would receive desks, badges, laptops, and publishing rights. In its subsequent update, OpenAI specified that outside evaluators may be brought into its offices for sensitive evaluations—a step the company said it has taken previously—though it has not finalized access terms or officially named evaluation partners.

The move follows OpenAI’s announcement last week that it would embed evaluators from Accenture under a contract valued at at least $1 billion over five years. In that announcement, OpenAI stated that funding for evaluations should ideally come from pooled or government sources because neither mechanism currently exists. The arrangement intersects with existing commercial ties: Accenture represents Anthropic's largest Claude Code deployment, and Accenture acquired Faculty, the London-based firm leading its evaluation work, in January.

The shift toward continuous evaluation coincides with European regulatory requirements. Under Article 55, providers of general-purpose AI models posing systemic risks have been required since August 2025 to perform adversarial testing to standardized protocols, alongside systemic risk assessments, cybersecurity protections, and prompt incident reporting. The European Union Agency for Cybersecurity (ENISA) already evaluates models, including those from OpenAI.

Safety governance has also spurred policy discussions around antitrust rules. Anthropic Chief Executive Dario Amodei requested an antitrust waiver in Washington to allow competitors to coordinate on safety measures, a proposal Altman endorsed. In the European Union, the European Commission ended individual exemptions in 2004, leaving companies to self-assess compliance under Article 101. The EU horizontal cooperation guidelines currently include no safety chapter, and no formal request has been made to add one.