OpenAI security staff reportedly asked three safety researchers to unlock their personal phones while investigating suspected leaks of confidential model-design information, according to an account attributed to The Wall Street Journal’s October 23 report.
The Journal’s original article was not available for verification in the source material reviewed for this article. The reported phone-access requests and scrutiny of Signal messages remain unverified here. Neither should be treated as proof that confidential information was leaked, that particular messages constituted misconduct, or that the subsequent firings were retaliation.
Earlier reporting and OpenAI’s public response document the underlying employment dispute. Tomek Korbak, Mikita Balesni and Jasmine Wang deny mishandling confidential information. OpenAI says its investigation established policy violations and explicitly denies firing them for raising safety concerns.
The dispute raises a practical question: how does a frontier AI lab authorize meaningful collaboration with independent evaluators while protecting sensitive research?
Phone-Access Requests Are Not the Same as Confirmed Searches
The account attributed to the Journal describes personal-phone access requests, scrutiny of Signal conversations and contacts with external safety researcher Ryan Greenblatt. Without the original report, the precise scope of those requests, what access investigators obtained and which communications they examined cannot be established here.
Asking an employee to unlock a phone does not, by itself, establish that investigators searched it. Even if an external conversation was examined, that would not establish that its contents violated company policy. Neither action alone would settle the reason for a dismissal.
The October 23 account concerns firings that had already been reported. TechCrunch’s October 8 report documented the three researchers’ dismissals and their open letter challenging OpenAI’s explanation. The potential news in the later account is the investigation process.
OpenAI and the Researchers Disagree About the Boundary
The researchers’ letter said they had not shared confidential company information and warned that their dismissals could make colleagues afraid to raise concerns or work with outside experts. They also denied involvement in a leak to The Information concerning model architectures that could make chain-of-thought reasoning harder to monitor.
They described external collaboration as part of their assigned safety work. TechCrunch reported their assertion that Balesni coordinated with leadership and removed sensitive details before sharing materials.
In its October 9 statement, OpenAI said its investigation uncovered a “significant breach of trust” beyond what the researchers described publicly. The company attributed the dismissals to sensitive-information policy violations and denied that they concerned safety criticism.
As CBS News reported, OpenAI also said it tolerates good-faith mistakes, encourages critical research discussions and does not terminate employees for raising concerns.
Neither account amounts to independently established findings about the underlying conduct. The public responses cited here do not identify enough of the allegedly mishandled material to determine which account is correct.
Supporting External Audits Does Not Define Disclosure Permissions
Both sides publicly support independent safety evaluation. Their disagreement concerns the boundaries of that work.
The researchers argue that outside expertise is necessary to understand and address emerging risks. OpenAI’s October 9 statement likewise described external collaboration as a core part of its safety work and said it was finalizing contracts with third-party assessors. Those arrangements were still described as work underway; the statement did not confirm their completion.
Support for collaboration does not establish permission to disclose every category of research information. A confidentiality obligation, in turn, does not explain how an employee should obtain authorization to discuss a safety concern with an approved evaluator. Assessing the firings would require evidence about both permission to maintain an external working relationship and permission to share particular information through it.
The researchers’ letter concerned preserving the ability to scrutinize model reasoning, and OpenAI publicly agreed that frontier-model monitorability requires an industry-wide commitment. Meaningful technical discussion can require detail. The operational challenge is to define which details an evaluator may receive, through which channels and with whose approval.
The Missing Evidence Concerns Authorization and Process
The public record leaves several questions unresolved: what information allegedly crossed a prohibited boundary, which rules governed it, and what authorization or guidance the researchers received.
The former employees’ warning about a chilling effect deserves attention, but it remains their assessment. The cited reporting does not demonstrate that OpenAI employees broadly stopped escalating concerns or that external audits were curtailed. OpenAI’s stated commitment to independent assessment also leaves unresolved whether the rules were clear or consistently applied in these cases.
External safety oversight needs understandable information-sharing permissions and usable escalation procedures. The dispute shows disagreement over those boundaries. The available evidence does not establish whether they were clear, followed or violated here, and does not support a finding of retaliation or misconduct.
