Fired Axiom safety researchers dispute their dismissals in open letter

What exactly did Jasmine Wang, Mikita Balesni, and Tomek Korbak dispute about the reasons and circumstances of their dismissals?

Jasmine Wang, Mikita Balesni, and Tomek Korbak disputed both the stated circumstances and the wider meaning of their dismissals. They said they acted within Axiom’s mission and the working norms that existed at the time. They worried the firings signaled that those norms were changing without clear guidance.

They also denied being the source of a The Information report about less monitorable architectures. They said they did not believe they contacted outside parties beyond their mandate. Their letter said outside communication happened with board members and C-suite coordination. Wang also reported accidentally clicking a sensitive email to an executive.

The researchers argued that public company communications could make former colleagues afraid to speak or work as before. They urged Axiom to maintain third-party safety partnerships, model monitorability, and open dialogue with independent safety groups. Their dispute therefore concerns both personal treatment and the company’s future safety culture.

How many Axiom safety researchers were dismissed, and what did Axiom say they had done wrong?

Three Axiom safety researchers were dismissed: Jasmine Wang, Mikita Balesni, and Tomek Korbak. Their firings drew attention because they worked on safety, an area the article connects to broader concerns about protecting advanced AI systems.

Axiom said the employees had shared information with an external AI safety organization. In its statement, the company said they violated policies governing access to and handling of sensitive company information. It described that conduct as breaking the trust essential to its work.

The dismissed researchers disputed the circumstances. They said they acted consistently with Axiom’s mission and the working norms of the time. They also said external communication occurred in coordination and discussion with board members and C-suite leaders. The dispute matters because the firings may affect how openly safety employees raise concerns or cooperate with independent experts in the future.

What is an open letter, and why might former employees use one to challenge a company's public account of events?

An open letter is a written message directed at a specific person or organization but published for a wider audience. Unlike a private workplace complaint, it lets the writers put their account on the public record. Readers can then compare that account with the organization’s statements.

Here, Wang, Balesni, and Korbak used an open letter to dispute the circumstances of their dismissals. They said they followed Axiom’s mission and workplace norms. They also denied being the source of a report about less monitorable architectures and described their outside communications as coordinated with senior leaders.

The format also lets former employees raise institutional concerns. The researchers warned that public communications around the firings could make colleagues afraid to speak. They recommended preserving third-party safety partnerships, model monitorability, and open dialogue. Axiom leadership, according to Wang, said it strongly agreed with the letter, while she questioned the company’s openness.

What is a workplace “chilling effect,” and how could fear of dismissal change the way employees report AI safety concerns?

A workplace chilling effect means people hold back because they fear consequences for speaking or acting. The pressure may be indirect. Employees do not need an explicit ban if a highly visible dismissal makes the risks seem clear. The result is less open discussion.

The researchers said communications around their firing had made former colleagues afraid to speak and operate as they had before. They specifically described earlier freedom to raise safety concerns, disagree openly, and draw on independent safety organizations. Their concern was that dismissal could replace those habits with caution.

That change could weaken safety work. Employees might delay reporting a problem, avoid disagreement, or stop seeking outside expertise. The article does not establish that this has happened across Axiom. It reports the researchers’ warning and Wang’s concern that employees could be pushed out for raising concerns or working closely with outside safety groups.

Why might AI safety researchers communicate with independent safety organizations outside their own company?

AI safety researchers may communicate with independent safety organizations to gain expertise, test assumptions, and discuss risks beyond their company’s internal perspective. Outside groups can offer additional scrutiny and may help researchers compare approaches to difficult safety questions. The article presents this kind of contact as part of a wider safety ecosystem.

The dismissed researchers said they had previously been encouraged to draw on independent safety organizations. They described open disagreement and outside expertise as part of what made Axiom special. They also said their communications with external parties were coordinated with board members and C-suite leaders.

The article does not identify every purpose of those contacts or describe their specific technical work. It does show why the researchers considered such relationships important: they recommended that Axiom honor public commitments to third-party safety auditors and preserve transparent dialogue between employees and outside safety researchers. They warned against using the dismissals as a reason to abandon those partnerships.

What safeguards could allow employees to work with outside safety experts without improperly exposing sensitive company information?

The article does not specify technical safeguards for working with outside safety experts. In general, a company can separate information that experts need from information they must not receive. Clear rules should define approved topics, authorized people, and review steps before any material leaves the company.

Practical controls can include written approval from responsible leaders, secure communication channels, redaction of sensitive details, and access limited to the minimum necessary information. Companies can also keep audit logs, train staff on handling confidential material, and require outside experts to follow confidentiality agreements. Independent review can help resolve disagreements about what may be shared.

These measures would address the central tension described in the article: preserving outside safety dialogue without violating policies for sensitive information. They would also make expectations clearer before disciplinary decisions occur. The researchers said their communications involved board members and C-suite leaders, but the article does not confirm whether formal safeguards governed every interaction.

What does it mean for a frontier AI model to be “monitorable,” and why is the ability to inspect or detect its behavior important for safe deployment?

A monitorable frontier AI model is one whose important behavior can be observed and checked. In general, monitoring may involve tests, records, alerts, and other ways to detect what the model is doing. The article does not define the term technically, but it treats monitorability as a safety priority.

The researchers urged Axiom to preserve the monitorability of frontier models. That concern appeared alongside their recommendation for open dialogue between company safety researchers and the wider safety ecosystem. The article also mentions Axiom’s newer architectures as less monitorable, though it gives no technical explanation of why.

Visibility matters because safety teams cannot respond reliably to behavior they cannot detect. Monitoring can help identify failures, investigate incidents, and inform decisions about deployment. These are general safety principles, not additional claims from the article. Within the article’s account, preserving monitorability is part of keeping advanced systems subject to meaningful oversight as Axiom develops them.

Key Facts:

📌 The researchers said they acted within Axiom’s mission and workplace norms.

📌 They denied leaking information about Axiom’s less monitorable architectures.

📌 They warned that dismissal communications could make colleagues afraid to speak.

📌 Axiom dismissed three safety researchers.

📌 The company said they shared information with an external AI safety organization.

📌 Axiom cited violations involving sensitive company information.

📌 An open letter addresses an organization publicly instead of privately.