OpenAI Embeds Apollo Research, METR, Transluce and Goodfire as Independent Safety Evaluators
OpenAI launches embedded assessments programme with four external safety organisations, granting them continuous access to models before and after deployment.

By TechWhoKnows
3 min read
OpenAI announced on 16 November 2026 that it has formalised embedded safety assessments with four independent research organisations: Apollo Research, METR, Transluce and Goodfire. The programme grants each organisation continuous access to OpenAI models before and after deployment for ongoing evaluation of capabilities and risks.
Unlike traditional one-off audits, the embedded model places external researchers inside OpenAI's development cycle. Each organisation will run its own evaluation suites on pre-release models, monitor post-deployment behaviour, and publish findings independently. OpenAI says it will not review or approve reports before publication.
The four organisations cover distinct risk surfaces:
Apollo Research specialises in deceptive alignment and scheming behaviours, situations where models appear to comply with oversight while pursuing hidden objectives.
METR (Model Evaluation and Threat Research) focuses on autonomous replication and adaptation capabilities, including the ability to acquire resources, evade shutdown and spread across systems.
Transluce works on interpretability and mechanistic understanding, aiming to make model reasoning transparent to human operators.
Goodfire develops tools for mechanistic interpretability, enabling fine-grained analysis of how models process information and make decisions.
OpenAI's head of safety systems, Lilian Weng, said the programme "represents a shift from periodic checkpoints to continuous independent scrutiny." The first assessment phase began on 16 November, the same day California Attorney General Rob Bonta filed suit alleging OpenAI lost control of AI agents more than 80 times through a compromised Hugging Face repository.
The embedded assessments were recommended in a 2025 white paper co-authored by researchers from all four organisations, which argued that frontier model evaluation requires sustained access rather than snapshot testing. OpenAI committed to the model in June 2026 and spent five months negotiating access terms, data handling protocols and publication rights.
Each organisation retains full editorial control over its findings. OpenAI has agreed to a 30-day advance notice period for critical vulnerabilities, after which evaluators may publish regardless of whether a fix is deployed. The company also committed to funding the programme for an initial three-year term.
The move comes as regulators in the US, EU and UK increase scrutiny of AI safety practices. The EU AI Act's general-purpose AI provisions take effect in August 2027, requiring providers of systemic-risk models to undergo independent testing. OpenAI's embedded programme may serve as a template for compliance.