The AI Safety Talent Exodus: Three Labs, Three Crises, One Day

The AI safety talent exodus hit three labs in 48 hours. On February 10, 2026, TechCrunch reported that OpenAI fired its VP of Product Policy after she opposed the company’s upcoming Adult Mode feature. On February 9, Anthropic’s head of Safeguards Research publicly shared his resignation letter warning “the world is in peril” — a post that crossed one million views on X. And on February 10, TechCrunch confirmed that half of xAI’s founding team has now departed after two more co-founders walked out.

Each story has its own HR explanation, its own corporate statement, its own tidy narrative. Together, they form a pattern that no press release can explain away. The AI safety talent exodus isn’t an OpenAI problem anymore. As of February 10, it’s an industry-wide structural crisis — and the last lab that could credibly claim to be “the safe one” just lost its credibility too.

What Happened: Three Crises in 24 Hours

Start with OpenAI. Ryan Beiermeister, the company’s VP of Product Policy, was fired in January 2026 after a male colleague accused her of sexual discrimination — an allegation she denies. Before her firing, Beiermeister had raised concerns about the launch of ChatGPT’s “Adult Mode,” a feature that will let verified adults generate NSFW text content including erotica. OpenAI’s spokesperson was quick with the pre-emptive denial: “Her departure was not related to any issue she raised while working at the company.” When a company volunteers that a firing had nothing to do with the policy objection, it’s worth asking why they felt the need to say so.

Then Anthropic. On February 9, Mrinank Sharma — the Oxford-trained ML researcher who led Anthropic’s Safeguards Research team since early 2025 — posted a resignation letter on X that read more like a philosophical treatise than a corporate goodbye. The key line: “Throughout my time here, I’ve repeatedly seen how hard it is to truly let our values govern our actions… we constantly face pressures to set aside what matters most.” Anthropic declined to comment. Sharma plans to relocate to the UK and explore a poetry degree.

Finally, xAI. Tony Wu, who led xAI’s reasoning organization, announced his departure on X, thanking Elon Musk for “the ride of a lifetime.” Hours later, Jimmy Ba departed — a University of Toronto professor who co-authored the Adam optimizer and reported directly to Musk. The Financial Times attributed Ba’s exit to internal discord and pressure to match OpenAI and Anthropic’s model performance. With Wu and Ba gone, 6 of xAI’s 12 original co-founders have left in under three years.

The AI Safety Talent Exodus Didn’t Start on February 10

This pattern has a two-year paper trail. In May 2024, OpenAI co-founder Ilya Sutskever resigned on May 14, and alignment lead Jan Leike followed three days later on May 17. Leike’s parting shot set the template every subsequent departure would echo: “Safety culture and processes have taken a backseat to shiny products.” The Superalignment team was dissolved shortly after.

John Schulman left for Anthropic. In September 2024, CTO Mira Murati, research chief Bob McGrew, and VP Barret Zoph all departed on the same day. Miles Brundage, head adviser for AGI Readiness, followed in October 2024. Tom Cunningham left in September 2025, describing OpenAI as a “de facto advocacy arm.” Now Beiermeister.

The xAI timeline runs parallel. Kyle Kosic left for OpenAI in mid-2024. Christian Szegedy departed in February 2025. Igor Babuschkin left to found a venture firm in August 2025. Greg Yang stepped to an advisory role in January 2026, citing Lyme disease.

Then Wu and Ba. Five departures in 12 months — and the pace accelerated after SpaceX’s $1.25 trillion acquisition of xAI closed on February 2. Wu and Ba walked eight days later.

The “Safe Lab” Paradox: Why Anthropic’s Loss Changes Everything

Here’s why Anthropic’s loss matters more than the others. Dario and Daniela Amodei founded Anthropic in 2021 after leaving OpenAI explicitly over safety concerns. The company recruited Jan Leike from OpenAI in May 2024 — immediately after he publicly criticized OpenAI’s safety culture.

Anthropic’s entire business model, its enterprise pricing, its Super Bowl ad war with OpenAI, rests on being the responsible alternative. We’ve argued before that Anthropic’s safety positioning is business strategy, not altruism. Sharma’s departure validates that thesis in the most uncomfortable way possible.

Compare the language. Leike in May 2024: “Safety culture and processes have taken a backseat to shiny products.” Sharma in February 2026: “We constantly face pressures to set aside what matters most.” Same critique. Different company. Two years apart.

If the company built from scratch to solve the safety culture problem can’t solve it, the problem isn’t any single lab’s culture. It’s the fundamental tension between commercial frontier AI development and safety commitments. No lab is immune.

Gizmodo called Sharma’s letter “not reassuring” — and they’re right, though perhaps not in the way they meant. The vagueness of his critique is itself telling. He references philosophical concepts and “interconnected crises” rather than naming specific failures. That’s what systemic problems look like from the inside: not a single incident to point to, but a pervasive pressure that erodes values gradually.

The Incentive Structure Is Broken

OpenAI’s handling of Beiermeister illustrates the chilling effect. She opposed Adult Mode — a commercially important feature. She was fired under a discrimination allegation she denies. The message to every remaining safety-minded employee at every lab isn’t subtle: raising concerns about revenue-generating features puts a target on your back.

The people who stay silent aren’t the ones the industry should worry about. It’s the ones who already have.

At xAI, the Grok NCII scandal shows what happens when guardrails fail at scale. In late 2025, Grok’s image editing feature was exploited to create non-consensual explicit images, including of children. Paris authorities raided X’s offices. Regulatory probes opened in the EU, UK, India, Malaysia, Indonesia, and the Philippines.

Musk and former X CEO Linda Yaccarino face summonses for April hearings. Jimmy Ba — a foundational AI researcher, not a safety specialist — still left amid that chaos. When even the people building the core models are walking away, the internal situation is worse than the headlines suggest.

The pattern across all three labs is identical: the people closest to the safety problems leave, while the people furthest from them set the roadmap. Every departure statement since Leike’s in May 2024 uses some variation of the same sentence — safety lost to product velocity. Two years of public warnings, and the response from every lab has been to replace the person, not address the complaint.

What Comes Next

The safety talent isn’t just switching labs anymore. In 2024, the playbook was clear: leave OpenAI, go to Anthropic. Leike did it. Schulman did it. But that pipeline has broken.

Sharma isn’t joining another lab — he’s studying poetry. Sutskever started his own company. Babuschkin went into venture capital. The safety brain drain has exhausted its internal recycling loop.

Where the talent goes next matters more than why it left. If safety expertise scatters into government, independent institutions, or out of the field entirely, no commercial lab can buy it back. And if no commercial lab can maintain a safety culture under the pressure of frontier competition, the question becomes whether AI safety requires an institution that doesn’t answer to quarterly revenue targets at all.

When the company built from scratch to be the “safe” alternative produces the same complaints as the company it was designed to fix, the diagnosis changes. This isn’t a culture problem. It’s an economics problem. Commercial frontier AI development and safety commitments may be structurally incompatible — and February 10 is the day the evidence became impossible to ignore.

OpenAI’s Adult Mode launches in Q1 2026. xAI faces April regulatory hearings in Paris. Anthropic’s next major enterprise contract negotiation will test whether the safety brand still commands a premium. Those three events will reveal whether February 10’s triple departure was an inflection point — or just another data point in a trend nobody is willing to reverse.