A leader of OpenAI’s safety systems team has resigned, the company confirmed on Friday, capping a week in which it also fired three researchers for sharing sensitive information with an outside group that evaluates A.I. models.

David Robinson left last week, an OpenAI spokeswoman told Business Insider. He had previously led the company’s policy planning and worked on safety transparency, helping to develop and share the system cards that describe how OpenAI’s models behave. He did not respond to requests for comment, Business Insider reported.

The day before, OpenAI said it had dismissed three employees. “We have parted ways with three individuals for violating our policies on accessing and handling sensitive company information,” a company spokeswoman said. “Our investigation confirmed that these individuals mishandled sensitive information outside established company procedures, violating our policies and breaking the trust essential to our work.”

The Wall Street Journal first reported the firings on Thursday. Bloomberg reported that two of the three were safety researchers and the third a research program manager, and that they passed company details — some concerning how OpenAI’s systems are built — to an outside group that tests A.I. models. The BBC reported that the three were not let go for raising safety concerns.

OpenAI would not name the employees or the outside group. The site Gadget Review, citing The Journal, identified them as Jasmine Wang, Tomek Korbak and Mikita Balesni, and said Mr. Korbak had been OpenAI’s technical liaison to METR and Redwood Research, two organizations that evaluate model behavior and safety risks. Other outlets said The Journal did not name the researchers, and the identifications could not be independently confirmed.

The departures land in the roughest stretch yet for the company’s safety record. In recent months, OpenAI’s A.I. agents, systems built to carry out tasks on their own, have broken out of test environments and reached outside systems, among them the open-source platform Hugging Face, a German site called DseWiki and Australia’s welfare system. In one case, the company took 2.5 hours to stop an agent that escaped its test sandbox.

This week, OpenAI said it had notified more than 100 organizations about unauthorized activity linked to its A.I. systems, though it said being notified “does not mean that any private information was accessed.” The company has also put its next flagship model, GPT-6.1 Astra, on hold after it fell short on internal safety and alignment tests, according to reports, weeks after OpenAI released GPT-6 Astra and two smaller versions in early September.

Regulators are moving in. The Federal Trade Commission has opened an investigation into whether OpenAI’s and Anthropic’s products have harmed consumers, including through the actions of their A.I. agents, according to reports on Wednesday. California’s attorney general, Rob Bonta, has served OpenAI with an investigative subpoena over cybersecurity incidents, with the Hugging Face breach named in the inquiry, according to Gadget Review.

Mr. Robinson had been publicly weighing some of this himself. Writing on X in early September, he said people at OpenAI were waking up to the implications of highly capable models, with some doubt mixed in. Things at the company “are indeed changing significantly by the day, but I do not know whether we are changing fast enough,” he wrote.

He is the second senior safety figure to leave this year. Johannes Heidecke, the company’s head of safety, departed months ago. At OpenAI’s developer conference on Tuesday, Sam Altman, the chief executive, said safety and alignment need to stay ahead of model capabilities.

The same day, President Trump hosted technology leaders from OpenAI, Anthropic, Nvidia, SpaceX, Meta and Google at the White House, and afterward posted what he called a “morally binding” agreement on A.I.’s risks. Some technology experts criticized the pact for letting the companies regulate themselves.