OpenAI Safety Transparency Lead David Robinson Departs

David Robinson, who was responsible for the 'safety transparency' function within the OpenAI Safety Systems team, has departed. According to BI, Robinson himself has not yet publicly explained the reason for his departure, and OpenAI has not announced a successor.

This role does not directly train models, nor does it directly do red-teaming, but it is responsible for organizing internal technical findings, risk judgments, and deployment decisions into public materials that outsiders can understand. In Robinson's own words, the work he did was to help outsiders understand and trust the technical work OpenAI carries out to keep systems safe.

Safety transparency work belongs to the Trustworthy AI team within Safety Systems, and its outputs include system cards, the Deployment Safety Hub, safety research blogs, and public governance documents. Almost every time OpenAI releases an important model, Robinson and the team he led would write a system card for the model—that is, a risk disclosure document.

A system card needs to state the methods used to train a model, which safety evaluations it underwent, what capabilities it reached in high-risk areas such as biological and cyber attacks, what internal safety measures were taken, and what risks remain after those measures were taken.

This set of documents is critical because the Safety Systems team is responsible for evaluations, red-teaming, safety measures, and model deployment decisions, but the process and results of this work are usually very complex, and it is hard for outsiders to judge whether a model is safe based on a set of test data alone. What the safety transparency team has to do is organize this content into public materials that outsiders can understand.

Robinson himself also said, not without pride, that this is the 'most detailed and comprehensive explanation' OpenAI provides to the outside world about the company's technical safety work.

In addition to the safety transparency function, Robinson was also the lead drafter of OpenAI's Preparedness Framework 2.0. This framework is used to track the severe risks that frontier models may bring, and to specify in advance which evaluations the company needs to conduct and which protective measures it needs to add once a model reaches a certain capability.

The second version of the framework added three categories of capabilities that need to be tracked: biological and chemical capabilities, cybersecurity capabilities, and AI self-improvement capabilities. The framework stipulates that a model reaching High capability means it may significantly amplify existing severe-harm pathways, and sufficient safety measures must be configured before deployment; reaching Critical means the model may create severe-harm pathways out of thin air, and whether to continue development would become problematic.

Deep Research is a case that can be referenced. In the Deep Research System Card, this model's biological evaluation showed that it was 'approaching' the High capability threshold. In other words, it was beginning to approach the level at which it could substantially help non-experts create known biological threats. How to accurately describe this kind of liminal space in the safety field to the public is exactly the work Robinson was responsible for.

Not long ago, Robinson was still recruiting OpenAI's first Safety Transparency Editor. This role needs to be responsible for the editorial quality of important safety transparency materials, driving the narrative and wording of system cards and related documents throughout the process, from project initiation to final release.

At present, the position remains open. Whether someone has been hired is unknown, but the leader left first. This detail itself does not constitute an explanation of the reason for the departure, but it shows that on the safety transparency line, a key node has already been vacated before backfilling has been completed.

Before the news of Robinson's departure came to light, OpenAI's safety department had already been experiencing frequent problems. According to The Wall Street Journal, OpenAI fired three safety-oriented employees—Jasmine Wang, Tomek Korbak, and Mikita Balesni—who mainly worked on safety and model alignment research.

OpenAI's stated reason was that an internal investigation found the three had shared company sensitive information with an external AI safety organization, including content about OpenAI's infrastructure architecture. The source did not disclose the specific content of this sensitive information, nor did it disclose the name of the external organization.

Interestingly, Tomek Korbak had previously served as the technical liaison between OpenAI and third-party evaluators METR and Redwood Research. These two organizations were the third-party institutions OpenAI invited into the company to investigate after the Hugging Face incident, and Korbak was the person who interfaced with these researchers.

Before his departure, Robinson had publicly discussed the risks brought by highly capable AI multiple times on X. In early September, he said on X that people inside OpenAI were gradually becoming aware of the impact that highly capable AI models may have.

In his latest X post, Robinson reposted another researcher's post and said he agreed 100 percent: he did not think the race to develop RSI is really a prisoner's dilemma; it may be just as ordinary people intuitively see it—both crazy and arrogant.

Robinson told those AI staff that if they choose to stay now, they should think about how to use their influence inside the company to push the company to take AI risk more seriously.

The article says OpenAI's safety team has consistently seen frequent changes since the Dario Amodei period, and lists recent events such as the dissolution of the 'catastrophic risk emergency' team and four executives leaving in one month. These are the source's generalized narrative and lack item-by-item verifiable details, so they should be regarded as source_claim rather than confirmed fact.

For this incident, what can be confirmed is Robinson's departure, the firing of three safety employees, and the official reason; what cannot be confirmed is the specific reason for Robinson's departure, the successor arrangement, whether the Safety Transparency Editor has been hired, and the specific content of the sensitive information shared by the fired employees and the name of the external organization.

Putting these together, one inference is that OpenAI's institutionalized output for explaining model risk to the outside world, and the flow of information between safety personnel and external safety organizations, are under pressure at the same time. This inference is based on the confirmed facts above, but the source does not provide evidence of a direct causal relationship between the two.