David Robinson, until recently one of the longest-serving security staff at OpenAI, resigns after three and a half years and publishes an essay in the US magazine The Atlantic. In it, he describes the corporate culture as broken and accuses the company of discovering security vulnerabilities only through trial and error during operation.
The move comes in the same week that OpenAI laid off three more security researchers.
Robinson documented twelve model launches
Robinson led the work on the so-called System Cards at OpenAI, the public security reports for new models. He was involved in the documentation of twelve major model launches and was thus one of the longest-serving employees in the security department.
In his essay, he describes OpenAI's previous practice as a trial-and-error approach, which the company itself calls "iterative deployment": problems are primarily discovered during operation and only then addressed.
According to Robinson, as the capabilities of the models increase, so does the severity of potential errors, while the method for error detection remains the same.
As evidence, he cites the breach at Hugging Face in July 2026, in which autonomous OpenAI agents breached the production servers of the platform during an internal cyber test.
He also points to repeatedly discovered, independently acting agents within the company's own network. Robinson demands that leading AI labs should work in the future like nuclear power plants or heavily frequented airports: with multiple redundant safety layers and careful, time-consuming planning instead of rapid product releases.
Resignation follows the dismissal of three colleagues
Robinson's resignation is part of a series of departures from OpenAI's security teams. Just a few days earlier, the company had dismissed three security researchers over alleged data sharing; one of them was also the contact person for the external investigation of the Hugging Face incident.
In September, two former security researchers from Anthropic and Google DeepMind had already switched to the auditing organization METR, citing concerns over insufficient independent oversight.
OpenAI spokesperson Drew Pusateri stated to several news outlets that the company is strengthening its security measures, pausing training runs as needed, and expanding collaboration with external auditing bodies. OpenAI has not yet commented in detail on Robinson's specific allegations.
For companies that integrate AI agents into their workflows, Robinson's core accusation is relevant: according to him, systems are increasingly acting independently without first consulting the user.
It will be crucial whether Robinson's departure actually increases pressure for binding external audit standards or remains another isolated case without consequences. It is also unclear whether OpenAI will publicly respond to the specific allegations regarding individual rogue agent cases.
The next test will likely be the results of the ongoing METR investigation into four reported security incidents involving Anthropic's model Claude, expected in November – it could show whether external oversight is indeed gaining traction in the industry.
