OpenAI's Firing Spree and What It Reveals About AI Safety

The Firing That Refuses to Stay Quiet

Mikita Balesni and Tomek Korbak are no longer at OpenAI, but they are finally speaking. Their first public interview since their dramatic dismissal marks a rare crack in the company’s carefully managed narrative — one that could reshape how regulators, competitors, and the public view the ongoing tension between speed and safety in frontier AI development. This is not the typical exit story where a pair of engineers fade into the background, quietly collecting stock options and LinkedIn recommendations. What distinguishes this moment is the specificity of the claims being made — and the institutional weight behind the people making them.

The researchers were part of OpenAI’s safety team, a group tasked with ensuring that increasingly powerful models don’t cause unintended harm. Their firing was swift and somewhat opaque, leaving questions that the company has yet to fully answer. Now, for the first time, they’re stepping outside to explain what happened — and what it means. The interview reveals a culture clash that has been building for years, one that extends far beyond any single organizational dispute and speaks to something far larger: the question of who controls the trajectory of artificial intelligence when the stakes are existential and the timelines are terrifyingly short.

What makes this story particularly striking is that Balesni and Korbak were not marginal figures on the safety team. They were deeply embedded in OpenAI’s internal safety architecture, involved in the kind of work that requires institutional memory and hard-won expertise. Their departure leaves a gap that isn’t easily filled, and their decision to speak publicly suggests they believe the public has a right to understand what went wrong — not just what happened to them personally, but what their dismissal signals about the direction of one of the most consequential technology companies on Earth.

Inside the Hugging Face Investigation

One of the most consequential details emerging from this interview involves METR’s investigation into the Hugging Face hack. Hugging Face, the open-source AI platform hosting millions of models, suffered a significant security breach that raised alarms across the AI community. The hack occurred during a period of heightened tension in the open-source ecosystem, where developers had already been wrestling with concerns about the provenance of training data, the exposure of proprietary model weights, and the broader question of whether open distribution was keeping pace with the capabilities of the models being shared.

METR, a prominent AI safety research organization with deep roots in the technical community, was brought in to investigate. Their findings, as described by Balesni and Korbak, pointed to vulnerabilities that were both technically serious and symbolically loaded. What the researchers are revealing is that OpenAI’s response to the incident — and to METR’s findings — has been deeply unsatisfactory from a safety perspective. The researchers say the company’s rebuttal “doesn’t add up,” suggesting either that OpenAI minimized the severity of the breach or that it failed to adequately address the vulnerabilities METR identified.

This matters because Hugging Face serves as infrastructure for the broader AI ecosystem. If a major platform hosting open-source models can be compromised, the implications extend far beyond OpenAI’s own closed systems. The hack exposed how interconnected and fragile the AI supply chain has become. Models trained on data scraped from the internet move through platforms like Hugging Face into research labs, startups, and government agencies around the world. A breach at that level doesn’t just affect one company — it threatens the integrity of an entire generation of AI systems.

The second-order effects of this situation are difficult to overstimate. When a leading lab dismisses external safety findings without adequate public justification, it creates a precedent that other companies will notice. The signal sent is clear: security concerns can be managed through corporate communication rather than through technical remediation. For an industry that already faces accusations of racing past its own safety protocols, that message is corrosive. It erodes trust not just in OpenAI but in the broader ecosystem of AI research and deployment.

The Safety Wars Inside OpenAI

The dismissal of Balesni and Korbak didn’t happen in a vacuum. It’s the latest chapter in an ongoing cultural struggle within OpenAI itself — a battle between teams prioritizing rapid advancement and those insisting on rigorous safety protocols. The internal fight over AI safety, as the researchers describe it, is still very much alive inside the company, even if some of its most vocal advocates have been removed. This is a pattern that has played out in various forms at other technology companies facing similar pressures, but the scale and visibility of OpenAI’s efforts make its internal dynamics especially consequential.

What Balesni and Korbak’s account adds is a sense of the human cost of this struggle. These are people who joined one of the most ambitious projects in the history of computing because they believed the work could be done responsibly. Their departure suggests that, at least in their view, the balance has shifted decisively away from that commitment. The safety team at OpenAI has, by all available accounts, been under sustained pressure for some time. Budget allocations, headcount decisions, and strategic priorities have all pointed in a direction that emphasizes capability development over safety assurance.

This dynamic reflects a broader tension across the AI industry. As models grow more capable, the questions become harder: Who decides when a system is safe enough to release? What happens when the people asking those questions are told to shut up? And what responsibility does a company like OpenAI owe to the public when the stakes involve systems that could reshape society? These are not questions that can be answered through press releases or strategic ambiguity. They require institutional mechanisms — independent oversight, transparent reporting, real consequences for unsafe deployment — and the current trajectory at OpenAI suggests those mechanisms remain underdeveloped at best.

OpenAI has positioned itself as a leader in responsible AI development, particularly after its controversial shift from a non-profit to a capped-profit structure. That shift was justified at the time as necessary to attract the capital required for the scale of ambition the company had articulated. But events like this — the firing of safety researchers, the handling of the Hugging Face investigation, the dismissive treatment of external scrutiny — challenge that positioning in tangible ways. The gap between rhetoric and practice is where trust is built or broken, and in this case, the gap is becoming harder to ignore.

Why This Matters Beyond Silicon Valley

The story of Balesni and Korbak’s dismissal is not just an internal company drama. It touches on questions that policymakers, regulators, and international bodies are already grappling with. The European Union’s AI Act, the frameworks being discussed in Congress, the deliberations of the UN’s advisory body on artificial intelligence — all of these efforts are shaped by events like the ones now unfolding at OpenAI. If a company as prominent and well-resourced as OpenAI cannot create an environment where safety concerns are taken seriously internally, it raises fundamental questions about whether external regulation can compensate for those failures.

If a company as prominent as OpenAI cannot contain dissent over safety concerns within its own walls, what does that signal about the governance of frontier AI more broadly? What safeguards exist when the primary mechanism for raising alarms — internal channels — appears to penalize rather than protect the alarm-raiser? These are not abstract concerns. The capabilities being built at OpenAI and comparable labs worldwide are outpacing the regulatory frameworks designed to oversee them. In that environment, internal safety voices become crucial early-warning systems. Removing or silencing them doesn’t just hurt the individuals involved — it degrades the system’s ability to detect and respond to real risks.

The geopolitical dimension adds another layer of urgency. The United States and China are locked in a competitive race that rewards speed above all else. Every milestone reached by one side creates pressure on the other to respond. In that context, safety considerations are easily framed as luxuries — things that can be addressed once the technology is mature enough, once the market position is secure enough, once the competitive threat has passed. But the trajectory of AI development suggests there may not be a “once” moment. The risks evolve as the capabilities evolve, and the window for meaningful intervention narrows with each new benchmark.

What Happens Next

Balesni and Korbak’s decision to go public suggests they believe the story deserves wider attention. Whether it gets that attention depends on how much weight the industry and the public are willing to give to their account — especially when it directly contradicts OpenAI’s version of events. The company has not issued a detailed rebuttal to their specific claims, which leaves the narrative in a space of competing testimonies. That kind of ambiguity tends to favor the people telling the more dramatic story, particularly in an era when institutional credibility is already fragile.

For now, the researchers have broken their silence. The interview provides a rare window into the internal dynamics of one of the world’s most influential AI labs. What they’ve said so far raises more questions than it answers, and the full implications of the Hugging Face hack and METR’s investigation are still unfolding. But the basic arc is already visible: a company struggling to reconcile its safety commitments with its competitive pressures, a team of researchers who believe they were pushed out for doing their jobs, and a broader industry that is watching closely to see what precedent will be set.

The debate over AI safety is not going away. It’s moving from lab benches and internal memo chains into public forums, podcast episodes, and regulatory hearings. The people who spent years working on these problems inside OpenAI are now speaking — and the industry would do well to listen. The question that remains is whether anyone in a position to make a difference actually will.