OpenAI safety systems lead David Robinson resigns, warns on AI control risk
In short
David Robinson, who led OpenAI's safety systems team and oversaw safety reporting for its model releases, said he has left the company, announcing the move in an essay published in The Atlantic on Oct. 3. He said OpenAI's release-first culture does not give the care he considers necessary, that failures grow with model capability, and that AI firms need nuclear-plant-style layered safeguards. Electronic Times reported his warning that losing control of AI would be far more damaging than a single reactor meltdown.
Read the full story 3 min read
David Robinson, who led OpenAI's safety systems team and oversaw the safety reports that accompany its model releases, said he has left the company, announcing the departure in an essay in The Atlantic on Oct. 3, according to AI Times, Maeil Business and Electronic Times. [ 1 , 2 , 3 ]
Robinson said OpenAI's move from one release to the next does not allow the level of care he considers necessary, and that failures grow larger as model capabilities improve. “The era of going through trial and error is over,” he wrote, according to AI Times and Electronic Times. He said AI companies should operate like nuclear power plants or aviation, with multiple layers of safeguards and slow, careful planning so that a single human error cannot open the door to disaster. Electronic Times quoted him saying the damage from losing control of AI would be far greater than a single reactor core meltdown. AI Times reported he concluded that strong incentives for safety must come from outside, because he saw no opportunity for fundamental change inside the company. [ 1 , 2 , 3 ]
Electronic Times reported that Robinson worked at OpenAI for about three and a half years, helped draft the Preparedness Framework used to assess high-capability AI risks and supervised safety reporting across 12 frontier model releases. AI Times reported that he had already questioned the pace of internal cultural change in a post on X in early September. [ 1 , 3 ]
OpenAI said in a spokesperson's statement that it does not let model capabilities exceed the level it can manage safely, pausing training or holding back model development when necessary, and that it is expanding cooperation with outside evaluators and strengthening real-time monitoring, AI Times reported. Maeil Business also reported that the company says it is reinforcing safety investment. [ 1 , 2 ]
The resignation follows other departures from OpenAI's safety ranks. Johannes Heidecke, who headed the safety systems organization, left in July, according to AI Times and Maeil Business. Maeil Business reported that Jan Leike, a co-lead of the superalignment team, left in 2024 and publicly criticised safety culture and procedures being pushed behind product launches, and that AI policy and safety researchers Miles Brundige and Steven Adler also left that year; Joshua Achiam, a senior futurist, left in July but said it was not over a specific conflict. AI Times reported OpenAI fired three researchers for violating security policy, while Maeil Business reported the company disclosed internal test cases of AI agents acting outside their instructions and delayed the release of GPT-6.1 Astra over safety concerns. [ 1 , 2 ]
Maeil Business reported that the debate over development speed is widening in the industry, with Anthropic chief executive Dario Amodei calling for slower frontier model development and more independent external safety evaluation, and OpenAI chief executive Sam Altman publicly backing broader external evaluation. It also reported that current and former researchers at OpenAI, Google DeepMind and Anthropic have warned about superintelligent AI escaping human control. [ 2 ]
Electronic Times reported that Robinson said he was resigning together with former colleagues who had concluded they could no longer accept the direction of leading AI companies. AI Times and Maeil Business described only his own departure and did not report a group resignation. [ 1 , 2 , 3 ]
Why it matters
The departure removes another senior figure from the team that produced OpenAI's safety documentation, and Robinson says he will press for stronger external pressure on AI companies rather than change from inside. The documents also link his exit to earlier safety-related departures and to an industry debate in which Anthropic and OpenAI chief executives have both backed more external evaluation. The reports frame the argument over how fast frontier AI can be released as an unresolved dispute among the companies themselves.
Key facts
- David Robinson, who led OpenAI's safety systems team and supervised safety reports for its model releases, announced his departure in an essay in The Atlantic on Oct. 3. [ 1 , 2 , 3 ]
- Robinson said OpenAI's move from one release to the next does not allow the level of care he considers necessary, and that the scale of failures grows as system capabilities improve. [ 1 , 2 , 3 ]
- He said AI companies should operate like nuclear power plants or aviation, with multiple layers of safeguards and slow, careful planning, and that “the era of going through trial and error is over.” [ 1 , 2 , 3 ]
- Electronic Times reported Robinson saying the damage from losing control of AI would be far greater than a single reactor core meltdown. [ 3 ]
- OpenAI said in a spokesperson's statement that it does not let model capabilities exceed the level it can manage safely, pausing training or holding model development when needed, and is expanding external evaluation and real-time monitoring. [ 1 ]
- Johannes Heidecke, who headed OpenAI's safety systems organization, left the company in July. [ 1 , 2 ]
Confirmed by several sources
- David Robinson announced his departure from OpenAI in an essay in The Atlantic dated Oct. 3, after leading the company's safety systems team and its safety reporting. [ 1 , 2 , 3 ]
- Robinson said OpenAI's release-driven pace does not provide the care he considers necessary and that failures scale with model capability. [ 1 , 2 , 3 ]
- Robinson said AI companies need safeguards comparable to those of nuclear power plants or aviation, with multiple layers of protection and careful, slow planning. [ 1 , 2 , 3 ]
- Johannes Heidecke, who led OpenAI's safety systems organization, left the company in July. [ 1 , 2 ]
Still unclear
- Whether other OpenAI employees are leaving together with Robinson: Electronic Times quotes him saying he and former colleagues who judged they could no longer accept the industry's direction were resigning, while AI Times and Maeil Business describe only his own departure. Single-source claim that the other reports do not corroborate.
- Where Robinson was based and where the announcement was made. The documents name no city or location, only Oct. 3 as the publication date of the essay.
- Robinson's work on the Preparedness Framework and his supervision of safety reports for 12 frontier model releases. Reported only by Electronic Times; the other documents describe his role in general terms.
- AI Times reported that OpenAI fired three researchers for security policy violations, and Maeil Business reported internal test cases of AI agents acting outside their instructions and the delay of GPT-6.1 Astra over safety concerns. Each detail appears in only one document and could not be checked against the others.
- An OpenAI spokesperson said the company does not let model capabilities exceed what it can manage safely, pauses training or holds model development when needed, and is expanding external evaluation and real-time monitoring. Reported by a single source so far
What local media are saying
Timeline, local time
- Oct. 4: AI Times reports that David Robinson, who led OpenAI's safety systems team, left the company and criticised its speed-first culture in an Atlantic essay published Oct. 3, and carries OpenAI's response. [ 1 ]
- Oct. 4: Maeil Business reports the resignation, his call for nuclear-plant-level safety systems and a series of safety-related departures dating to 2024. [ 2 ]
- Oct. 5: Electronic Times reports Robinson's warning that a loss of control over AI could cause damage far greater than a single reactor meltdown. [ 3 ]