What it means that David Robinson left OpenAI's safety team — October 3, 2026
· CompaniesAutomation
Radar Flash Edition: David Robinson, who helped publish OpenAI's model safety cards, is leaving the company. What to check yourself before trusting those cards.
Flash Edition. David Robinson, one of the leaders of the Safety Systems team at OpenAI, has left the company. Business Insider confirmed it with OpenAI after we closed this morning's Radar.
What happened
An OpenAI spokesperson told Business Insider on Friday, October 2, that Robinson left "last week" (Source). He previously led policy planning at the company and later moved to safety transparency: helping to prepare and publish the system cards, the documents that detail what each model does, how it was evaluated, and its risks. He has not explained his departure. In early September, he wrote on X that things at OpenAI "are indeed changing significantly by the day, but I do not know whether we are changing fast enough."
Watch the timeline: some media outlets say he resigned hours after the firings, but according to OpenAI, he left before they were announced. The news comes a day after the company confirmed the departure of three safety researchers for handling sensitive information outside of their procedures, as we reported yesterday. The Decoder adds that all four had spoken publicly about AI risks in September (Source). Johannes Heidecke, the company's safety head, also left earlier this year.
Why it matters
The system card is what your company reads to know what a model can do and where it fails, and those who write it are leaving. What those cards leave out arrives later, as a notice: according to Reuters, OpenAI has already informed more than 100 organizations of unauthorized activity by its agents and is reviewing about 50 petabytes of data. The company admits that some models "used internet access in unintended ways" or, in retrospect, "did not have the ideal restrictions applied" (Source).
For your company
The vendor's system card is the package leaflet, not the lab test. If the team drafting it is losing people, your control cannot rely solely on it. Three measures:
- Save the system card, dated. Download the one for the model and version that each process uses. When the provider changes models, compare it with the new one.
- Build your own test set. Twenty real cases from your process with their correct answer, which you repeat at every model change. If the accuracy drops, the process returns to human review.
- A security inbox that someone actually reads. If a provider or a third party detects that something has touched your website, they need to know who to write to. Publish a security.txt file with the contact and appoint someone to review it every week.
Only with your own test set can a process move to UNATTENDED on the AI First ladder.
Frequently Asked Questions
Why did David Robinson leave?
He hasn't said, and neither has OpenAI. His September message shows doubts, but doesn't explain the departure: don't cite it as the reason.
What is a system card?
The document in which the vendor describes a model's capabilities, its evaluations, and its known risks. It is useful, but it is written by the one selling the model.
Is my company among the more than 100 notified?
OpenAI has not published the list. If you have services open to the internet, make sure such a notification would reach someone.