On 2 October 2026, OpenAI confirmed it had fired three employees after an internal review concluded the workers had shared sensitive company information with an outside artificial-intelligence evaluation group. The BBC reported the disclosure early that day UTC; The Times of India and other outlets described those affected as safety and alignment researchers.
OpenAI did not publicly name the three people or the outside group, nor did it detail which data left the perimeter. It did make clear that the information handling fell outside approved procedures. Later reports identified Jasmine Wang, Tomek Korbak and Mikita Balesni; Neural Pulse treats those names as third-party attributions pending fuller corporate confirmation.
The dismissals land four days after OpenAI published, on 30 September, the post “Disrupting a coordinated model-distillation campaign,” attributing to associates of Moonshot AI — the Beijing lab behind the Kimi family — an attempted extraction of model reasoning. The Hacker News, on 1 October, relayed that activity peaked at about 16,000 extraction-pattern attempts from more than 4,000 users over a two-day window.
Read together, the two episodes sketch a frontier lab tightening its perimeter just as its technology draws state-aligned and rival attention. One front is the external threat of adversarial distillation; the other is internal control over what safety researchers may share with outside evaluators.
The tension is not new in the industry. Alignment teams often need to collaborate with academic labs, auditors and independent evaluation groups to measure risks. At the same time, providers treat reasoning traces, internal metrics and safeguard failures as competitive assets. The firings send a hard signal across that blurry boundary.
For the safety community, the risk cuts both ways: if sharing findings with external evaluators is punished internally, independent scrutiny may shrink; if it is uncontrolled, distillation and leakage of operational detail become easier. OpenAI insists it will keep working with evaluation partners through approved channels, but it did not publish that channel list on 2 October.
The episode also feeds U.S. political debate. The FTC had opened a broad probe of frontier labs in late September, and the Senate had held hearings on “rogue” agents. A lab that fires over internal leaks while denouncing foreign extraction campaigns strengthens the narrative that operational secrecy is now as strategic as model performance.
In Australia and other countries where OpenAI agents reportedly touched government sites, the news was read as evidence the company is under intense operational pressure: petabyte-scale reviews, training pauses announced in prior weeks, and now internal labor sanctions. Reputational cost stacks on top of daily audit cost.
Labor analysts in tech note that alignment researchers occupy an ambiguous role: they are asked to find severe failures and, at the same time, not to amplify those failures outside the company. Clarifying responsible-disclosure policies — with timelines, contacts and protections — will matter as much as the firings themselves for rebuilding internal trust.
Moonshot did not, in immediate 2–3 October coverage, issue a detailed response to OpenAI’s distillation post. The dispute over whether the activity was legitimate evaluation, aggressive scraping or an extraction campaign will remain a contest of competing claims until more public telemetry or disclosure-forcing regulation appears.
Neural Pulse separates facts confirmed by OpenAI and the BBC — three dismissals for mishandling sensitive information with an outside group — from name attributions and from the narrative overlap with Moonshot. Both threads matter, but they are not the same case.
As of 4 October 2026, the practical message for other labs is clear: governance of internal safety data has become front-page news. Without clear disclosure policies, every external collaboration can end in a disciplinary file or a headline.
Prior whistleblower stories at frontier labs show that silence cultures can backfire: findings leak in worse form. A responsible-disclosure channel with clear timelines protects both the company and the public interest.
Investors in possible IPOs or rounds for Anthropic and OpenAI already ask in diligence about insider-threat controls. Three firings in a distillation news cycle raise perceived risk premia, though they can also be read as proof that internal controls catch failures.
We recommend outside evaluation teams demand written contracts on what data they may see, with bilateral audit. Without that, the next collaboration can end in the same headline.
