When Safety Teams Walk: Governing Internal Dissent at Frontier Labs

On October 9, 2026, OpenAI dismissed three safety researchers, Tomek Korbak, Jasmine Wang, and Mikita Balesni, according to an AP report. The company stated that the dismissals followed an investigation which found the researchers had violated clear policies regarding the handling of sensitive information. OpenAI maintained that the decision was not related to safety concerns or the researchers speaking out, asserting that the company cannot perform its work without a high degree of trust.

The Collision of Proprietary Interests and Safety Oversight

The dispute involving Korbak, Wang, and Balesni highlights a fundamental tension in the governance of frontier AI labs: the overlap between intellectual property protection and the requirements of safety auditing. While OpenAI has emphasized the need for trust and the adherence to sensitive information protocols, the researchers involved have offered a different interpretation of the events. Tomek Korbak stated on X that his dismissal was linked to his communication with METR, an outside AI evaluation nonprofit that was engaged to investigate a previous incident involving Hugging Face. Korbak wrote that he believes the three were fired for prioritizing safety over the near-term interests of OpenAI as a corporation. OpenAI disputes that account.

This tension is not merely a matter of individual conduct, but a structural conflict inherent to the current model of AI development. In this model, frontier labs possess highly sensitive, proprietary datasets and model architectures. To maintain a competitive advantage, these companies implement strict controls on the dissemination of information. However, effective safety auditing often requires that researchers share findings or internal assessments with external evaluators to ensure transparency and independent verification. When the boundaries of "sensitive information" are defined by the corporation, the ability of safety teams to interact with the broader scientific community becomes subject to the discretion of management.

The friction between a lab's need for information security and the safety community's need for transparency is a central governance challenge for the next decade of AI development.

Key Insight

The Role of External Evaluators in Internal Governance

The involvement of METR in the investigation of the Hugging Face incident underscores the increasing reliance on third-party entities to validate AI safety. Mikita Balesni's work involved "cross-company work" intended to preserve the ability to monitor AI systems. According to Balesni, he took precautions to remove sensitive details from any materials shared externally. OpenAI says its investigation found the researchers violated policies on handling sensitive information. The researchers say they fear the dismissals chill the open disagreement about AI risks that the company's culture has encouraged.

The core of the dispute is unresolved. OpenAI says the researchers violated policies on handling sensitive information; Korbak says he was told it concerned how he communicated with METR, which he says was his job. If the former is true, then the dismissals represent a standard enforcement of corporate security protocols. If the latter is true, the dismissals may serve as a deterrent to researchers who believe that their duty to public safety outweighs the non-disclosure agreements of their employer. This creates a significant ambiguity for the industry, as the outcomes of such investigations are not independently corroborated by third parties.

Analyzing the Divergent Interpretations of Corporate Dismissal

To understand the implications of these events, one must examine the two primary scenarios that emerge from the documented facts. Each scenario offers a different view of how governance should function in a high-stakes technological environment.

Scenario A: The Institutional Compliance Model

In this scenario, the actions taken by OpenAI were a necessary enforcement of internal controls. As the company stated, the work requires a high degree of trust. If researchers are permitted to share sensitive internal data with external nonprofits, even with the intention of protecting safety, the integrity of the lab's proprietary information could be compromised. Under this model, the dismissal of researchers who violate communication protocols is essential to maintaining the organizational structure required to build and deploy frontier models.

Scenario B: The Transparency Deficit Model

In this scenario, the dismissals represent a prioritization of commercial interests over safety transparency. If Korbak's assertion is correct, that the dismissals were a response to prioritizing safety over near-term corporate interests, then the governance structure of the lab is designed to favor speed and secrecy over the ability to undergo rigorous external audit. This model suggests that as AI models grow more powerful, incentives may shift toward limiting internal safety discussions.

Implications for Industry Stakeholders

The dispute at OpenAI provides several indicators for different classes of observers. For executives within the AI industry, the primary concern is the management of talent and the stability of research culture. If researchers perceive their work is subject to dismissal for external collaboration, the industry may see experts move to academia.

For investors, the risk is centered on the unpredictability of governance. A lab that cannot maintain strict control over its information may face intellectual property risks, but a lab that suppresses safety concerns may face massive regulatory backlash or catastrophic model failures. This creates a bifurcated risk profile where the very mechanisms used to protect the company's assets (such as NDAs and strict communication policies) may simultaneously increase its long-term liability.

For policymakers and regulators, the OpenAI case raises the question of whether current oversight mechanisms are adequate. If regulators are to ensure that AI safety is being taken seriously, they cannot rely solely on the internal investigations of the companies they are regulating. The lack of independent corroboration makes it difficult to determine if the dismissals were security-based or intended to limit discussion.

Future Indicators for AI Governance

The resolution of the tension between safety and security will likely be signaled by several key developments in the coming months. Stakeholders should monitor the following indicators to assess whether the industry is moving toward a model of transparent oversight or one of centralized control:

• Verification of Internal Investigations: The emergence of independent, third-party audits that can corroborate or refute the claims made by companies like OpenAI regarding the reasons for personnel dismissals.
• The Status of External Evaluation Partnerships: Whether frontier labs establish formalized, legally protected frameworks for interacting with organizations like METR, or if communication with external evaluators remains a high-risk activity for employees.
• Legislative Mandates on Disclosure: The introduction of regulations that require labs to report dismissals of safety-critical personnel to a neutral third-party regulator.

Currently, it remains unknown whether the actions at OpenAI were an isolated incident of policy enforcement or a symptom of a broader industry trend toward the prioritization of commercial speed over external safety verification. The lack of independent corroboration regarding the investigation's findings leaves the core question of motive, whether it was a matter of security or a matter of interest, unresolved.

Sources: abcnews.com.