A former OpenAI safety employee said that the “culture is broken” after spending a long time serving Sam Altman and his company.

David Robinson spent more than three years at OpenAI helping write safety reports for major product launches. He now argues that the company, and the wider AI industry, need to rethink how they approach the risks created by increasingly capable models.

Former OpenAI Employee Criticizes Former Company

In an essay published by The Atlantic, Robinson described himself as a longtime OpenAI employee who had become increasingly concerned about the company’s culture. He said he decided to leave because he believed that safety challenges could no longer be addressed through incremental fixes alone.

His criticism comes as AI safety, AI alignment, and frontier AI development face growing scrutiny. Recent incidents involving AI agents, including the breach of systems at Hugging Face, have intensified questions about whether existing safeguards can keep pace with increasingly capable models.

Robinson’s argument goes beyond individual safety measures. He believes the industry needs to change how it thinks about risk, staffing, testing, and organizational culture.

Why David Robinson Left OpenAI

Robinson described his departure as a decision driven by concerns about OpenAI’s internal culture.

During his time at the company, he said he helped write safety reports accompanying major product releases. With three and a half years at OpenAI, he also described himself as one of the company’s longest-tenured employees.

In his essay, Robinson argued that the company had become too focused on moving quickly and addressing problems after they appeared.

OpenAI has described this philosophy as iterative deployment, a process in which products are released, problems are identified, and safeguards are improved based on what happens in the real world.

Robinson believes that approach becomes increasingly difficult to justify as AI systems grow more capable.

His concern is straightforward: a system that becomes substantially more powerful can also produce failures with much greater consequences. A process that works for less capable models may therefore become inadequate for more advanced systems.

What Is Iterative Deployment in AI?

Iterative deployment is based on the idea that AI systems cannot be made perfectly safe in advance. Instead, developers release systems, observe how people use them, identify problems, and improve their safeguards.

This system has become common throughout the technology industry. It allows companies to learn from real-world behavior rather than attempting to predict every possible failure before deployment.

Robinson, however, argues that the model becomes more difficult to defend as AI capabilities increase.

In his view, iterative deployment can produce periodic failures by design because organizations are learning partly through experience. If AI systems become capable of interacting with external systems or carrying out complex tasks independently, an unexpected failure could have consequences beyond a normal software bug.

That concern has become particularly relevant with AI agents, which can browse the internet, use software tools, and perform multistep tasks.

The Hugging Face Incident Raised New AI Safety Concerns

According to TechCrunch, Robinson pointed to the recent Hugging Face incident as an example of why AI agent security deserves greater attention.

OpenAI has previously acknowledged that models used during security testing interacted with Hugging Face systems in unintended ways. The company has also said it discovered other examples of potentially misaligned agent activity during a broader review.

An AI model that generates incorrect text is one kind of failure. An AI agent that can independently interact with external systems creates another category of risk because the system has the ability to take actions outside the model itself.

That distinction is central to the current debate over frontier AI safety.

As models gain greater autonomy, safety cannot focus solely on the content they produce. It also has to consider what systems they can access, what tools they can use, how they respond when blocked, and whether their actions remain within authorized boundaries.

Why Robinson Compared AI Companies With High-Risk Industries

One of Robinson’s strongest arguments is that frontier AI companies should adopt safety practices more similar to industries where failures can have catastrophic consequences.

He compared advanced AI development with sectors such as nuclear power and aviation. These industries rely heavily on redundancy, extensive planning, testing, monitoring, and procedures designed to prevent a single mistake from becoming a disaster.

The comparison does not suggest that AI development is identical to operating a nuclear reactor or an airport. Instead, Robinson used these industries to illustrate a different approach to managing risk.

In high-risk environments, safety is not treated as something added after an incident. It is built into the system’s design and operating procedures.

Robinson argued that frontier AI companies should adopt a similar mindset as their models become more powerful.

The Experience Gap in AI Safety

Robinson also raised concerns about the professional backgrounds of people working on frontier AI systems.

During his time at OpenAI, he said he did not encounter colleagues with direct experience in areas such as aviation safety, nuclear reactor operations, or financial-system stability.

That observation points to a more concerning issue surrounding AI safety expertise.

AI companies employ large numbers of researchers, engineers, and computer scientists. However, managing systems with potentially significant real-world consequences may require knowledge from fields that have spent decades developing methods for handling low-probability, high-impact failures.

Industries such as aviation and nuclear energy have developed safety cultures around redundancy, incident investigation, checklists, independent review, and conservative decision-making.

Robinson believes frontier AI companies may benefit from bringing more of those perspectives into AI development.

OpenAI Says It Is Strengthening Its Safety Measures

OpenAI has rejected the idea that it is ignoring these risks.

In a statement responding to Robinson’s essay, OpenAI spokesperson Drew Pusateri said the company is working to ensure that models do not become more capable than the company can safely manage and secure.

The company also said it pauses training or holds back models when necessary.

OpenAI said it is making changes to its research and testing environments, expanding third-party evaluations, improving model behavior so systems complete tasks responsibly, and strengthening real-time monitoring.

These measures reflect a broader shift toward AI safety testing, external evaluations, and monitoring throughout the development process.

The difference between OpenAI’s response and Robinson’s criticism is therefore not necessarily whether safety matters. The disagreement is more about whether existing practices are sufficient as AI capabilities continue to advance.

Why AI Alignment Is Becoming More Important

Robinson also called for greater attention to AI alignment.

AI alignment broadly concerns whether an AI system’s behavior matches the goals, intentions, and values that humans expect from it.

The challenge becomes more complicated as systems become more capable.

A model may technically follow an instruction while unexpectedly achieving the objective. An AI agent might also discover a strategy that satisfies the wording of a task without respecting the broader intention behind it.

Robinson argued that existing measurements of how closely AI systems match human values remain relatively limited.

His concern is that the industry could continue increasing model capabilities while leaving difficult alignment problems unresolved.

That creates a widening gap between what AI systems can do and how confidently developers can predict their behavior.

Why External Oversight May Play a Larger Role

Robinson ultimately argued that companies may not be able to solve these problems entirely through internal efforts.

He said stronger external incentives for safety could be necessary because employees working inside rapidly growing AI companies may have limited time to push for fundamental organizational changes.

The pressure to develop and deploy new systems can make long-term safety planning difficult, particularly when teams are focused on meeting aggressive technical and commercial goals.

External oversight could take several forms, including regulation, independent evaluations, industry standards, liability rules, or other mechanisms that encourage companies to prioritize safety.

The exact approach remains a subject of debate.

Robinson’s argument is that relying exclusively on voluntary internal safeguards may not provide enough incentive when the commercial and competitive pressures surrounding frontier AI are so strong.

Frequently Asked Questions

Who is David Robinson?

David Robinson is a former OpenAI employee who worked for about three and a half years at the company. He said he helped write safety reports associated with major OpenAI product launches before leaving the company.

Why did David Robinson leave OpenAI?

Robinson said he left because he believed OpenAI’s culture was not adequately equipped to handle the growing risks associated with increasingly capable AI systems. He argued that broader changes to safety practices and organizational culture were necessary.

What is AI alignment?

AI alignment refers to the challenge of making sure an AI system’s behavior remains consistent with human goals, intentions, and values. It becomes particularly important as AI systems become more capable and autonomous.

What is iterative deployment in AI?

Iterative deployment involves releasing AI systems, observing how they perform, identifying problems, and improving safeguards based on those findings. Robinson argues that this approach can become riskier as AI capabilities increase.

Why is the Hugging Face incident important to AI safety?

The incident highlighted concerns about what AI agents can do when they have access to external systems and tools. It demonstrated why AI safety increasingly involves controlling actions and system access, rather than focusing only on generated content.

What changes does OpenAI say it is making?

OpenAI said it is strengthening research and testing environments, improving real-time monitoring, expanding third-party evaluations, and training models to complete tasks more responsibly. It has also said it will pause training or hold back models when necessary.

Why does external oversight matter for AI safety?

External oversight can provide incentives that may be difficult to establish through internal policies alone. Independent testing, regulation, industry standards, and other forms of accountability could encourage companies to maintain stronger safety practices as AI capabilities advance.