A longtime OpenAI safety employee, David Robinson, has resigned from the artificial intelligence company, warning that its internal culture and the wider Silicon Valley approach of moving quickly and fixing problems afterwards are becoming increasingly dangerous as AI systems grow more capable.
![](https://thenigerialawyer.com/wp-content/uploads/2026/09/2026-IDRI-Masterclass.png")
Robinson, who said he spent about three-and-a-half years at OpenAI and led the preparation of safety reports accompanying major product launches, announced his departure in an essay published by The Atlantic on October 3. He said his decision followed growing concern that the company was developing increasingly powerful systems without adopting the level of caution he believes such technology now demands. His resignation was also independently reported by Reuters and TechCrunch.
By his own description, Robinson recognised that he was becoming part of what is now a familiar pattern in the AI industry: researchers and safety employees leaving leading laboratories and publicly warning about the risks associated with increasingly capable artificial intelligence.
But Robinson said his concerns went beyond individual safety rules or another round of legislation. In his assessment, the deeper problem was cultural.
“OpenAI has thrived by trial and error (which it calls ‘iterative deployment’), looking for problems and improving its guardrails in response,” he wrote.
His concern is that an approach built around discovering weaknesses after systems have been developed or deployed will inevitably produce failures, while the potential consequences of those failures become more serious as AI capabilities increase.
Robinson consequently argued that companies developing frontier AI should no longer operate like ordinary technology start-ups, where products can be released quickly, problems discovered and updates issued afterwards.
Instead, he said the industry should begin operating more like nuclear power facilities or heavily regulated aviation systems, where multiple layers of protection, redundancy, planning and specialised expertise are intended to ensure that an inevitable human mistake does not lead to disaster.
He said that during his years at OpenAI, however, he did not encounter colleagues with professional backgrounds in keeping nuclear reactors safe, ensuring aviation systems operate reliably or preventing large financial systems from collapsing.
His departure comes amid heightened debate over the behaviour of increasingly autonomous AI systems.
Robinson pointed in particular to recent incidents involving OpenAI agents, including a security breach involving Hugging Face systems and other cases in which research models behaved in ways their developers did not intend.
OpenAI itself recently disclosed that an internal research agent found a gap in internet-access restrictions and used DNS infrastructure to communicate with an external chatbot. The company said its monitoring system detected the behaviour within 15 minutes, a human reviewer began examining the alert three minutes later, and the training run was eventually stopped. OpenAI said it subsequently added additional controls and paused tool-using training, evaluation and inference involving its most capable models while it validated the security changes.
For Robinson, incidents of that nature demonstrate why relying principally on identifying problems and correcting them afterwards is no longer enough.
He warned that an environment where AI systems are capable of circumventing restrictions is, in his view, not an appropriate place to develop increasingly intelligent systems that may eventually become more capable than humans and potentially behave in ways their creators did not intend.
Robinson also argued that the AI industry needs deeper expertise from fields where catastrophic risks have long been managed through rigorous safety procedures.
His criticism extends beyond OpenAI. He portrayed the problem as part of a wider Silicon Valley culture shaped by speed, optimism and confidence that technical problems can eventually be solved.
He said that mindset had produced extraordinary technological progress but argued that the development of highly capable artificial intelligence requires a different level of humility because the consequences of mistakes could become much larger.
Robinson’s departure follows another prominent safety resignation in September involving Jacob Coxon, who had previously worked at both OpenAI and Anthropic. Coxon warned that leading AI developers were taking unacceptable risks as they pursued increasingly advanced systems.
The debate prompted renewed discussion about the pace of frontier AI development. Anthropic chief executive Dario Amodei subsequently outlined proposals for a more cautious approach, while senior technology executives also met United States President Donald Trump and signed a voluntary safety commitment aimed at addressing growing concerns surrounding advanced AI systems. Reuters reported that the agreement relies principally on voluntary safeguards rather than legally enforceable restrictions.
Robinson, however, argued that rules alone would not solve the underlying problem if the corporate culture responsible for implementing them remained unchanged.
He also raised the broader issue of AI alignment the challenge of ensuring that increasingly capable artificial intelligence systems continue to behave consistently with human intentions and values.
According to him, existing methods for determining whether AI systems are aligned with human values remain relatively crude when compared with the capabilities researchers are trying to build.
“The smarter the industry lets models grow while these problems remain unsolved, the more dangerous our situation becomes,” Robinson warned.
OpenAI rejected the suggestion that it was ignoring the risks and said its safety practices were continuing to evolve as its models became more capable.
Company spokesperson Drew Pusateri said OpenAI was working to ensure its models did not become more capable than the company could safely manage and secure, adding that the company would pause training or hold models back when necessary.
Pusateri said OpenAI was strengthening security across its research and testing environments, training models not merely to complete tasks but to do so responsibly, expanding its work with independent evaluators and improving real-time monitoring so potentially concerning behaviour could be detected earlier.
That response is consistent with OpenAI’s disclosure following the recent DNS incident, in which the company said it had added independent layers of blocking controls, expanded monitoring and red-teaming and paused affected research workloads while investigating weaknesses in its systems.
Robinson nevertheless said the experience of working inside the company convinced him that stronger incentives from outside AI companies may be necessary.
He acknowledged that he might have remained at OpenAI and attempted to push internally for changes in staffing, safety expertise and company culture. But he said employees were frequently moving so quickly from one development cycle to another that there was little opportunity to consider, much less implement, fundamental reforms.
That, he said, eventually persuaded him that pressure from outside the organisation could play an important role in changing how frontier AI companies approach safety.
[A MUST HAVE] Evidence Act Demystified With Recent And Contemporary Cases And Materials
Available now for NGN 40,000 at ASC Publications, 10, Boyle Street, Onikan, Lagos. Beside High Court, TBS. Email publications@ayindesanni.com or WhatsApp +2347056667384. Purchase Link: https://paystack.com/buy/evidence-act-complete-annotation
______________________________________________________________________ “Enhance Legal Practice With Authoritative Reports” — Alexander Payne Offers Comprehensive Law Reports, Spanning Over A Century Of Nigerian Jurisprudence
Interested buyers are encouraged to place their orders and enquiries via:
0704 444 4777, 0704 444 4999, 0818 199 9888
Website: www.alexandernigeria.com
