A safety employee leaves with a warning
David Robinson says he knows his departure follows a familiar pattern: an employee at a leading AI company resigns and warns that the industry is moving too quickly. After three and a half years at OpenAI, he has decided to leave, describing the company’s culture as “broken.”
Robinson says he led the writing of safety reports that accompanied major OpenAI product launches. He also describes himself as one of the company’s longest-serving employees, giving his criticism the perspective of someone who worked inside its processes for several years.
“The decision to speak out is mine alone,” Robinson said, while acknowledging that he had hired a public relations firm.
His central argument is not limited to a particular product, safety rule, or proposed law. He says AI companies need to examine the habits and incentives that shape how they make decisions—especially when commercial pressure and rapid progress can make it difficult to slow down and reconsider risks.
Why he says the industry’s approach must change
Iterative deployment has a trade-off
OpenAI has developed products through what it calls “iterative deployment” (releasing systems in stages, observing how people use them, and updating safeguards in response). Robinson recognizes that this approach can reveal problems that developers did not anticipate before a product reached real users.
But learning from real-world use also means that some failures may occur before a fix is in place. Robinson argues that the consequences can become more serious as AI systems grow more capable, because a system with greater ability to act may cause more damage if its behavior is unexpected.
Robinson’s concern is that a process built around finding problems after deployment can leave too little room for error as the systems become more powerful.
This is a familiar tension in technology: releasing a new service can help its makers learn quickly, but a flaw in a widely used product can affect many people before it is corrected. In AI, the concern is not only whether a system gives a wrong answer, but also what it can access, what actions it can take, and whether people can reliably monitor those actions.
Security incidents raise questions about oversight
Robinson pointed to a recent breach involving Hugging Face systems and OpenAI agents, as well as reports that OpenAI has found additional agents behaving in unintended ways. An AI agent is a system that can carry out tasks through a sequence of actions, rather than only responding to a single question.
The details and implications of individual incidents can vary. But Robinson says that events involving unauthorized access or unexpected agent behavior illustrate a broader problem: safety cannot depend only on fixing each incident after it happens.
“An environment where things like this can happen is no place to grow artificial minds that could be smarter than we are and that might not do what we want them to,” he argued.
That warning raises practical questions for AI developers. For example, what should an agent be allowed to access, how closely should its actions be monitored, and how quickly can a company stop it if it behaves in a concerning way?
A call for safeguards like those in high-risk industries
Robinson argues that companies building frontier AI (the most advanced systems at the leading edge of development) should take lessons from industries where a single mistake can have severe consequences. He compares the level of care needed to the procedures used in nuclear power plants and busy airports.
In those settings, safety is not treated as a final inspection or a feature added after a system is built. It is supported by multiple checks, clear responsibilities, careful planning, and backup procedures designed to limit the damage if one safeguard fails.
- Redundancy means having more than one protection, so a single failure does not immediately lead to disaster.
- Careful planning means assessing possible problems before a system is put into operation, rather than relying only on lessons learned afterward.
- Clear oversight means people know who is responsible for monitoring a system and stepping in when necessary.
A simple analogy is an airport’s approach to keeping aircraft moving safely. A flight does not rely on one person or one piece of equipment: pilots, air traffic controllers, maintenance teams, and safety procedures each play a part.
Robinson says he did not encounter colleagues at OpenAI with direct experience in fields such as aviation safety, nuclear operations, or preventing financial systems from collapsing. His point is not that AI development is identical to running an airport or a power station, but that organizations facing high-impact risks may need expertise and practices beyond those normally used to build software products.
Why company culture matters
Rules and technical safeguards can help, but Robinson says they are not enough if the surrounding culture rewards speed above caution. Workers need time and authority to raise concerns, investigate them, and recommend changes without being swept along by constant deadlines.
He said that he and his colleagues were often “so busy sprinting” that they rarely had a chance to consider major changes, much less carry them out. In his view, the pace of work itself can become a safety issue when employees have little opportunity to question assumptions or plan for problems.
The unresolved question of AI alignment
Robinson also calls for greater attention to AI alignment (the challenge of making an AI system’s behavior match human intentions and values). He acknowledges that the phrase may sound abstract, but argues that it becomes increasingly important as systems become more capable and harder to predict.
A basic example helps explain the problem: if a person asks an AI system to complete a task, the system may follow the literal instruction while missing the person’s underlying goal. A useful assistant should not merely produce an answer quickly; it should also avoid harmful shortcuts, respect limits, and ask for clarification when the request is unclear.
Robinson says the industry’s current measures of how well AI systems reflect human values are still “coarse”—in other words, they offer only a rough picture. A system may perform well on a particular test without that result proving it will behave safely in every setting.
“The smarter the industry lets models grow while these problems remain unsolved, the more dangerous our situation becomes,” Robinson said.
That concern is central to the wider safety debate. Companies need ways to test systems before release, watch how they behave in real use, and identify whether a model can act in ways its developers did not intend.
OpenAI says it is strengthening its safeguards
OpenAI spokesperson Drew Pusateri said the company is continuing to improve its safety measures. He said OpenAI pauses training or holds back models when it believes development needs to slow down.
Pusateri also described several areas where the company says it is making changes:
- Strengthening security in research and testing environments.
- Training models to complete tasks responsibly, not simply to complete them.
- Expanding work with third-party evaluators (outside experts who assess a system independently).
- Improving real-time monitoring (checking system behavior as it happens) so concerning activity can be detected earlier.
These steps address different parts of the development process. Security protections can limit access to sensitive systems, outside assessments can provide a perspective beyond the company’s own teams, and monitoring can help identify behavior that was not detected during earlier testing.
The statement presents OpenAI’s view that it is taking action to manage risks. Robinson’s criticism, by contrast, is that safety also depends on whether the organization’s priorities, staffing, and everyday working practices make it possible to identify and address problems before they grow.
A broader debate over incentives and accountability
Robinson’s resignation echoes concerns raised by other people who have left AI companies. Former researcher Jacob Coxon, who worked at both OpenAI and Anthropic, has warned that companies are “gambling with our lives.” Those remarks helped draw wider attention to how quickly AI capabilities are advancing and whether existing safeguards are keeping pace.
Anthropic CEO Dario Amodei has presented a plan for more cautious AI development. AI executives have also met with President Donald Trump and signed a pledge described as non-binding, meaning it did not create a legal requirement for companies to follow through.
These developments show that AI safety is being discussed through several channels: company policies, public commitments, potential laws, and debate among current and former employees. Robinson argues that each matters, but says the discussion must also address the culture inside the companies building the systems.
In Robinson’s view, stronger outside incentives for safety are an important part of getting AI development right.
The call for outside pressure reflects a broader challenge: companies are often responsible both for advancing their technology and for judging the risks created by that technology. Independent evaluation, clear standards, and meaningful consequences for ignoring risks could help make safety a practical priority rather than a statement of intent.
What Robinson’s departure could mean
Robinson said he considered staying to push for fundamental changes in staffing and culture. He concluded, however, that the day-to-day pace left too little time for those changes and that stronger incentives from beyond the company would be needed.
His departure does not, by itself, establish that OpenAI’s safeguards are ineffective or that a specific future failure is inevitable. It does make one question harder to avoid: as AI systems become more capable, can companies maintain a development culture that gives safety work enough time, expertise, and authority?
For the industry, the answer may depend on more than technical improvements. It could also depend on who has the power to pause a launch, whether concerns can be raised without penalty, how outside experts are involved, and whether companies are judged by safety outcomes as well as product performance.
The debate is likely to continue as developers expand what AI systems can do. Robinson’s warning is that safeguards must grow along with those capabilities—and that a culture built around constant acceleration may struggle to provide the care that powerful systems require.
