OpenAI and three former safety researchers agree on one consequential event and dispute its meaning. Jasmine Wang, Tomek Korbak and Mikita Balesni were dismissed after an internal investigation. OpenAI says they violated policies for accessing or handling sensitive information and committed a significant breach of trust. The researchers say they acted within their safety responsibilities and that abrupt, unclear discipline could make colleagues afraid to raise concerns or work with external evaluators.[1][2][3]
The complete investigative record, exact information shared, written approvals, warnings and policy text have not been published. That prevents an outside article from deciding the employment dispute or concluding that no rulebook exists. The narrower governance question is whether permissions for external safety work are clear, usable and open to review.
Known, disputed and unresolved
| Category | What the public record supports | How it should be reported |
|---|---|---|
| Established event | The three named researchers were fired after an internal investigation. | State as fact.Source 1Source 2 |
| OpenAI position | The conduct violated sensitive-information policies and was not retaliation for raising safety concerns. | Attribute directly to OpenAI.Source 1Source 3 |
| Researchers’ position | Their outside communication was connected to assigned safety work and the dismissals create a chilling effect. | Attribute to their letter and public accounts.Source 2Source 3 |
| Unresolved facts | The exact material, permissions, policy clauses, management directions and investigative evidence are undisclosed. | Do not infer or fill the gap.Source 1Source 3 |
| Area of overlap | Both sides publicly support continued safety work and external oversight in principle. | Separate agreement on goals from disagreement over conduct.Source 3 |
External evaluation creates a real information problem
An independent evaluator cannot assess a frontier system with no access to incidents, transcripts, model behaviour or technical context. A company cannot hand over every internal record without limits: disclosures may contain personal data, security-sensitive details, model vulnerabilities, proprietary research or information about uninvolved employees. The safety channel must therefore provide enough access to challenge the company while defining what can leave, through which route and under whose authority.
Korbak says his work involved communication with the external evaluator METR after an unprecedented agent incident. Balesni’s public account concerns cross-company work on model monitorability. Wang says she accidentally opened a sensitive executive email through access previously delegated for recruiting, reported it within minutes and had already asked for the access to be removed. These are their accounts; the public record does not independently adjudicate them.[3][2]
A minimum external-safety rulebook
| Measure | Value | Context |
|---|---|---|
| Mandate | Named purpose and evaluator | State which programme, incident or model behaviour the outside party is authorised to assess. |
| Information classes | Public, evaluator-confidential, restricted and prohibited | Researchers should know the boundary before the next novel incident. |
| Approved channel | Named recipients, storage, encryption and retention | Safety-critical collaboration should not depend on informal personal practice. |
| Redaction and review | Fast pre-disclosure review with an emergency route | A process must work quickly enough for live risk, not only routine publication. |
| Audit trail | Material accessed, transformed, approved and transmitted | Protects the organisation and the researcher when later accounts conflict. |
| Appeal and non-retaliation | Independent review of disputed good-faith disclosures | Distinguishes deliberate leakage from a safety escalation made under ambiguous rules. |
Ambiguity is itself a safety failure mode
A policy can be formally strict and operationally vague. “Work closely with outside experts, but never mishandle sensitive information” transfers the most difficult judgement to an individual researcher. If norms are developed in real time and later enforced as if every boundary had been settled in advance, even justified discipline can look arbitrary. The resulting fear can suppress the early, incomplete warnings that a safety organisation needs most.
The opposite boundary matters too. Calling a disclosure “safety work” cannot automatically authorise publication of confidential model details, exploit paths or personal information. Durable oversight needs institutional access rather than heroic rule-breaking: pre-authorised evaluators, case-specific data rooms, narrow confidentiality duties, rapid escalation and a record that can be reviewed by an independent body.
- Publish the categories of information external evaluators may receive and the process for exceptions.
- Give safety employees a written mandate, named approver and emergency escalation route.
- Record policy changes with effective dates rather than applying a newly clarified rule retrospectively without explanation.
- Provide independent appeal for discipline tied to safety disclosures or evaluator collaboration.
- Report aggregated use of the channel: cases opened, access granted, material withheld and reasons.
- Preserve evaluator independence by preventing the laboratory from controlling the final assessment.
The next useful disclosure is not another broad assurance that safety remains important. It is the procedure that will govern the next incident: who may contact an external evaluator, what evidence they may share, how urgency changes the approval path, who reviews a disputed decision and how employees are protected when they use the channel in good faith. Frontier oversight becomes credible when the people closest to a risk know exactly how to bring it outside the institution without guessing where loyalty ends and misconduct begins.
Sources & further reading
- OpenAI says it fired three researchers for violating sensitive information policy Reuters · 9 Oct 2026 · Accessed 10 Oct 2026
- OpenAI fires three safety researchers in dispute over AI risks Associated Press · 9 Oct 2026 · Accessed 10 Oct 2026
- Fired OpenAI safety researchers dispute misconduct claims and warn of chilling effect TechCrunch · 8 Oct 2026 · Accessed 10 Oct 2026