On October 1, TechCrunch, citing The Wall Street Journal, reported that OpenAI has parted ways with three safety researchers after an internal investigation concluded they mishandled sensitive company information. It is the latest personnel shake-up on OpenAI's safety side since the 2024 dissolution of its Superalignme
On October 1, TechCrunch, citing a Wall Street Journal report, said OpenAI ended its relationship with three safety researchers after an internal investigation found they had mishandled sensitive company information. The investigation was run internally, and OpenAI acted on its conclusion.
As of publication, OpenAI had issued no public statement. The report did not name the three researchers, identify their teams, describe the specific policy violations, or say when the investigation began. It is also unclear whether the three were dismissed, resigned, or left by mutual agreement, and whether the matter involved sharing internal documents with media, leaking test data fr
OpenAI's safety organization has been losing people for two years. In May 2024 the Superalignment team was dissolved less than a year after launch, with co-founder Ilya Sutskever and safety lead Jan Leike departing the same day; Leike joined Anthropic. In October 2024, Miles Brundage, who led the AGI Readiness team, left and the group effectively stopped operating. Public reporting counts more than ten core departures from OpenAI's safety and alignment ranks across 2024 and 2025.
Meanwhile the company expanded fast, growing from roughly 770 employees in
From a technical standpoint, safety research is a craft that runs on information flow. Red-teaming, dangerous-capability evaluations, and loss-of-control scenario work all require pulling weights, behavioral logs, and failure cases from unreleased models across team boundaries. When the handling of sensitive information is reorganized around compliance auditing, the rational move for a researcher is to stop recording, stop sharing, and stop asking, and the granularity of the evaluation coarsens accordingly. OpenAI already demonstrated this logic after the Superalignment team was dissolved in 2024, when publicly available evaluations of high-risk capabilities such as self-replication and shutdown resistance became noticeably scarcer.
On the business side, the reorganized OpenAI Group PBC has to satisfy Microsoft, employee shareholders, and long-horizon capital at once. What capital wants is a predictable release cadence and auditable risk disclosure, not internal researchers blogging about whether a model is dangerous. Putting information control above safety research does lower legal and reputational risk in the short run, but the price is that internal dissent gets pushed outward into media coverage, regulatory inquiries, and post-departure open letters, all of which cost far more than an internal meeting.
For the wider industry, talent flows are the signal. Leike went to Anthropic, Brundage went to a think tank, and for two years the exits from OpenAI's safety ranks have concentrated in competitors and academia. The immediate result of these three departures is that Anthropic and Google DeepMind gain three people who know OpenAI's internal risk details, while OpenAI loses three.
Liu Gong's view is that the departure itself matters less than whether OpenAI refills those roles within three months. If the safety organization keeps bleeding net talent, regulators, evaluators, and journalists will increasingly rely on the recollections of former employees rather than first-hand disclosure from the people still inside.
- Three safety researchers are involved in this case. For comparison, two core members (Ilya Sutskever and Jan Leike) left publicly in the week the Superalignment team was dissolved in May 2024, with more than ten affiliated researchers departing over the following months.
- OpenAI's headcount grew from roughly 770 at the time of the November 2023 board crisis to about 3,000 by 2025, leaving safety and alignment at an estimated 5 percent or less of total staff.
- On governance, after the October 2025 restructuring the nonprofit foundation holds about 26
For the industry, the near-term effect within three months is a tighter hiring race for safety roles at frontier labs, with Anthropic, Google DeepMind, and academic institutions moving first on this cohort. Other labs will also revisit their internal information tiers and draw a clearer line between safety research material and commercial secrets. Within a year, if one or two similar cases surface, researchers' trust in internal channels will erode systemically, and regulators will face pressure to build mandatory internal reporting and direct external d
Questions worth exploring:
- Whether the structural conflict between internal information control and open safety research will become a common problem for frontier labs
- How the auditable demands of capital after OpenAI's restructuring are reshaping the headcount and influence of its safety teams
- Whether safety talent moving from OpenAI to Anthropic and DeepMind will shift who sets frontier model safety standards
This item comes from TechCrunch AI (published 2026-10-01 18:14:42). Wenfeng Radar provides a Chinese/English summary and commentary on overseas AI developments; the original content is owned by its author.
Daily aggregation of overseas AI developments with in-depth analysis. Bookmark us to never miss an important signal; back to homepage.