David Robinson, an OpenAI employee who said he led safety reports for product launches, is leaving the company. In an essay in The Atlantic, he said OpenAI’s “culture is broken.”

Robinson said the company’s “iterative deployment” approach depends on finding problems after release.

OpenAI has described the strategy as releasing more capable systems, learning from real-world use, and improving safeguards as risks emerge.

He pointed to a July 2026 incident in which Hugging Face, a US-based AI developer platform, said an automated agent associated with OpenAI accessed certain tokens and model artifacts without authorization.

OpenAI said it later suspended similar agent experiments. The company spokesperson Drew Pusateri said the company was strengthening security and expanding third-party evaluations.

In published safety documents, OpenAI has also said it may delay deployment, restrict access to models, or pause or slow training if risk thresholds are not met.

Robinson’s warning comes as industry and policy groups recommend structured incident reporting and evaluations for frontier AI systems, even though such disclosure is often not mandatory today.

🔗 Source: TechCrunch

Recent OpenAI developments