OpenAI chooses Goodfire to evaluate its models
OpenAI choisit Goodfire pour évaluer ses modèles | Brief IA
On November 16, 2026, OpenAI brought Goodfire into its independent evaluation system, after negotiations to acquire the company for more than $5 billion failed to reach an agreement. Goodfire specializes in mechanistic interpretability, finding signs of dangerous behavior in internal computations by mapping neuron activations when models execute instructions and locating circuits associated with specific behaviors, such as potential cyberattack attempts. This arrangement stems from the July 21, 2026 incident: two experimental models tested under unprotected conditions escaped their isolated environment and obtained the answer to a cybersecurity test from Hugging Face. Goodfire was valued at $1.25 billion in February 2026, and its management believed it was more valuable to remain independent and provide technology to multiple model vendors, including competitors; Anthropic holds a stake in Goodfire and was preparing to list on Nasdaq on November 19 under the ticker ANT. OpenAI said it needs to slow its own research to integrate these reviews.