Quick Answer
OpenAI says it fired three safety researchers for violating sensitive-information policies, while the researchers say the dismissals could chill internal safety work and outside evaluation. The dispute rests on competing accounts, not publicly detailed allegations. Readers should treat both claims as unresolved, review the researchers’ letter and OpenAI’s policy, and watch for fuller evidence or an independent account.
Key Takeaways
- Three former OpenAI researchers published a public letter on October 8, 2026.
- OpenAI said on October 9 that it had parted ways with the researchers after an investigation.
- The company said the researchers violated policies on handling sensitive information.
- The researchers asked OpenAI to protect independent evaluation and frontier-model monitorability.
- Neither side has publicly provided the detailed evidence needed to resolve the dispute.
What happened between OpenAI and the 3 safety researchers?
OpenAI fired Tomek Korbak, Jasmine Wang, and Mikita Balesni in the week before October 8, 2026, when the three researchers published a four-page public letter addressed to OpenAI’s Safety and Security Committee, Safety Advisory Group, and Mission Advisory Council. The letter is titled “OpenAI cannot make AI safe on its own,” and it argues that internal safety work needs meaningful outside scrutiny.
The former researchers said the dismissals and subsequent communications could discourage internal safety debate and collaboration with external evaluators. Their concern is not limited to their own employment. The letter frames the firings as a test of whether researchers can raise concerns and work with independent organizations without fearing retaliation or unclear disciplinary standards.
OpenAI’s public response came the following day. The company said it had parted ways with the three researchers after an investigation found violations of “clear policies on handling sensitive information.” The company also said the dismissals were not about safety concerns or speaking out, according to AP’s account of OpenAI’s response.
The central issue remains unresolved because the public record contains competing descriptions but not the underlying evidence. Readers should distinguish between OpenAI’s stated policy rationale and the researchers’ allegation that the actions could affect safety oversight more broadly.
What did the former OpenAI researchers say in their letter?
The three former OpenAI researchers asked the company to uphold commitments in 3 areas: embedding independent evaluators in safety work, preserving the monitorability of frontier-model reasoning, and clarifying procedures for researchers who work with third parties. The requests appear in the October 8 public letter, which sets out the researchers’ account directly.
Independent evaluators are outside organizations that assess AI systems rather than relying solely on the company developing the systems. The researchers argue that external evaluation can provide a separate view of model risks, safety controls, and whether internal claims hold up under scrutiny. This matters because an AI developer cannot fully establish public confidence by assessing its own work alone.
The letter also focuses on monitorability, meaning the ability to examine and understand important aspects of how advanced models reason or behave. The researchers’ argument is that monitorability can support safety research by making it easier to identify concerning behavior before a model is deployed or used at greater scale.
The requests do not establish that OpenAI failed to meet those commitments. The letter instead asks OpenAI to state clearer protections and processes. For readers following the company’s broader governance record, the dispute arrives while public attention on AI oversight continues to grow, including questions raised in the FTC’s AI agent investigation.
How did OpenAI defend the firings?
OpenAI defended the firings by saying the three researchers violated clear policies governing sensitive information. The company said it had investigated the matter and described the situation as a significant breach of trust, according to Reuters’ report on OpenAI’s October 9 statement.
OpenAI also said the employment actions were not about safety concerns or speaking out. That distinction is important because an organization can argue that it supports employees raising concerns while also enforcing limits on how potentially sensitive information is shared. The public statements available so far do not explain the specific material involved or identify the precise conduct that OpenAI concluded violated its policies.
| Public claim | Source of claim | What remains unclear |
|---|---|---|
| The researchers were fired after an investigation. | OpenAI’s October 9 statement | The detailed findings and evidence from the investigation |
| Sensitive-information policies were violated. | OpenAI’s October 9 statement | Which information and which policy provisions were involved |
| The firings could chill safety debate and external evaluation. | Korbak, Wang, and Balesni’s October 8 letter | How future OpenAI safety practices will change, if at all |
| Communication with METR was part of Korbak’s job. | Korbak’s account to AP | OpenAI’s detailed account of that communication |
The practical limitation is that neither a company statement nor a public letter substitutes for a complete evidentiary record. OpenAI has not publicly provided specific allegations, and the former researchers’ letter does not independently establish the facts behind the company’s investigation.
Why does METR matter in the OpenAI safety dispute?
METR matters because Tomek Korbak said he was told verbally that his firing concerned how he communicated with METR, an independent AI-evaluation nonprofit. Korbak said communicating with METR was part of his job, while OpenAI has not publicly described the specific communication at issue.
Independent evaluation is one way AI companies can test systems with researchers outside their own management structure. External organizations may examine a model’s behavior, assess safety controls, or investigate incidents using methods that differ from the developer’s internal processes. The benefit is additional scrutiny, although independent evaluators also need clear rules for access, confidentiality, and responsible handling of sensitive information.
The disagreement illustrates why written procedures matter. A researcher working with an outside evaluator needs to know what can be shared, who approves the exchange, how the work is documented, and what channel exists for raising concerns. Ambiguous rules can create conflict even when both the company and researcher describe their goals as safety-related.
OpenAI’s dispute with the researchers also follows other public scrutiny of its security and governance practices. Readers can compare the separation between claims and verified conduct in OpenAI’s response to a reasoning-extraction campaign, where the company described a specific security action rather than an internal employment dispute.
What does OpenAI’s raising-concerns policy say?
OpenAI’s January 12, 2026, raising-concerns policy says personnel should report suspected misconduct. The policy explicitly lists theft, privacy misuse, and AI-safety policy violations among examples of concerns that can be reported. OpenAI’s published policy establishes that the company formally recognizes AI-safety policy violations as reportable issues.
The policy is relevant because OpenAI’s statement says the dismissals were not about raising safety concerns. A reporting policy can provide a channel for employees to flag suspected problems, but it does not by itself answer separate questions about sharing information with third parties or whether a specific disclosure complied with confidentiality requirements.
OpenAI’s policy supports raising concerns, but the policy does not publicly resolve whether the three researchers’ conduct complied with sensitive-information rules. That distinction is the reason the current dispute cannot be settled from the policy language alone.
The most sensible interpretation is that both protections need clear boundaries. Employees need a reliable method to report potential misconduct, and organizations need rules that protect confidential information. A credible process explains how those obligations interact before a disagreement becomes a disciplinary case.
What changes are the researchers asking OpenAI to make?
The former researchers are asking OpenAI to preserve and clarify safeguards rather than describing a single technical fix. Their letter calls for independent evaluators to be embedded in relevant work, for frontier-model reasoning to remain monitorable, and for procedures governing third-party collaboration to be made clearer.
Embedding independent evaluators would mean giving external experts a defined role in assessing relevant AI safety work. Preserving monitorability would mean treating insight into advanced model behavior as a safety priority. Clearer third-party procedures would define what researchers can share, how approvals work, and how conflicts are escalated.
Those changes could make expectations easier to understand for both employees and external organizations. At the same time, the letter does not specify exactly how OpenAI should balance wider access for evaluators with the company’s concerns about sensitive information. That implementation question is important because safety research can involve material that organizations believe needs safeguards.
OpenAI has not publicly committed to the requests in the letter. Readers should therefore view the proposals as the former researchers’ stated recommendations, not as announced changes to OpenAI policy or governance.
How should readers interpret the OpenAI safety researchers firing dispute?
The OpenAI safety researchers firing dispute should be interpreted as an unresolved governance conflict, not as proof of either side’s full account. OpenAI says an investigation found policy violations involving sensitive information. The three former researchers say the firings and communications could weaken open safety debate and outside evaluation.
The disagreement matters because safety work depends on trust in several directions: between employees and management, between companies and independent evaluators, and between AI developers and the public. A company can have formal reporting policies and still face questions about whether workers understand the practical boundaries around external collaboration.
For most readers, the key evidence to watch is more specific information. OpenAI could explain the relevant policy provisions and the nature of the alleged violation without necessarily disclosing sensitive material. The former researchers could provide further documentation about the processes they believe should have protected their work.
Until then, the practical response is to avoid treating the dispute as settled. The available statements show a serious disagreement about AI safety governance, but they do not provide enough public detail to determine whether OpenAI’s disciplinary action or the researchers’ characterization is more complete.
FAQ
Did OpenAI fire three safety researchers?
Yes, OpenAI said it had parted ways with Tomek Korbak, Jasmine Wang, and Mikita Balesni after an investigation. The three researchers said in their October 8 letter that they had been fired during the prior week.
Why did OpenAI say it fired the researchers?
OpenAI said the researchers violated clear policies on handling sensitive information. OpenAI has not publicly provided detailed allegations or identified the specific information involved.
What did the researchers say about the firings?
The former OpenAI researchers said the dismissals and ensuing communications could chill internal safety debate and work with outside evaluators. Their letter asks OpenAI to protect independent evaluation and clarify procedures for third-party collaboration.
What is METR’s connection to the dispute?
METR is an independent AI-evaluation nonprofit that Korbak said he communicated with as part of his job. Korbak said he was told verbally that the communication with METR was connected to his firing, while OpenAI has not publicly detailed its account.
Has OpenAI publicly changed its safety policy after the letter?
No, OpenAI has not publicly announced a policy change in response to the October 8 letter. The public record currently includes the researchers’ requests, OpenAI’s stated reason for the dismissals, and the company’s existing raising-concerns policy.