Why 6 AI Safety Researchers Quit in One Week

Six people left in one week

Back view of a researcher leaving an AI company office

It happened in exactly 6 days, from February 9 to 14, 2026. Anthropic's Head of Safeguards Research left with a letter saying "the world is in peril." An OpenAI researcher quit after publishing a New York Times op-ed that "OpenAI is repeating Facebook's mistakes." At xAI, two co-founders quit within 48 hours. OpenAI's VP of Product Policy was fired after opposing adult mode.

People whose job was to make AI safe decided they could no longer do that job. Not one or two. Three companies, simultaneously, within a week. CNN reported the situation as "AI's gatekeepers are leaving and sounding the alarm."

Anthropic: "The world is in peril" letter

On February 9, Mrinank Sharma posted an open letter on X. He led Anthropic's Safeguards Research team. Defense systems preventing AI models from helping with bioweapon creation, teams researching AI sycophancy, teams building internal transparency mechanisms. The person directing all of this was leaving.

The core sentence of the letter is this: "The world is in peril. Not just from AI or bioweapons. Because of a series of interlinked crises." He wrote that he had "repeatedly witnessed how difficult it is to act according to values" inside Anthropic. Both within the organization and in society at large.

Sharma announced he would return to the UK to focus on writing, poetry, and community activities. A researcher from the frontlines of AI safety leaving to write poetry. PC Gamer described it as an "epic vaguepost," but the vagueness itself was the message. There were things he couldn't say explicitly.

AI safety system image with flowing digital code background

Sharma wasn't alone. Three more left Anthropic during the same period. R&D engineer Harsh Mehta and senior AI researcher Behnam Neyshabur announced on X in early February they were leaving to "start something new." Both praised Anthropic's talent pool and culture, yet chose to leave. AI safety researcher Dylan Scandinaro was more direct. He left Anthropic to join OpenAI as head of the Preparedness team. His judgment about where the frontlines of safety research lie had changed.

These departures happened as Anthropic closed a $30 billion funding round at a$ 38 billion valuation. The moment the company's valuation hit an all-time high, the people responsible for safety were walking out.

OpenAI: Repeating Facebook's mistakes

On February 11, Zoe Hitzig published an op-ed in the New York Times. The title was "OpenAI Is Making the Mistakes Facebook Made. I Quit."

Hitzig is an economist and poet who worked at OpenAI for 2 years. Her resignation was announced the very day OpenAI started testing ads in ChatGPT. Her first sentence on X: "I resigned from OpenAI on Monday. The same day, OpenAI started testing ads in ChatGPT."

Hitzig's point wasn't simply "ads are bad." It was a warning about ChatGPT's particularity. Users tell ChatGPT about health issues, relationship problems, religious faith, financial situations. It's different from Google search. Search is about finding information, but ChatGPT conversations are about exposing private thoughts. OpenAI holds the most detailed record of private human thought ever assembled, Hitzig wrote.

The Facebook analogy was specific. Facebook also promised to protect user data in the early days. But as the ad model became the core revenue source, engagement optimization came to dominate all decisions. Algorithms amplified rage and extremes not from malice but structural inevitability. Hitzig warned that ChatGPT would follow the same path. Once responses start being tuned to advertiser interests, the AI assistant stops being a user tool and becomes an advertiser channel.

OpenAI did not issue an official comment on Hitzig's resignation. They only stated the ad testing was in a "limited and exploratory phase." But Silicon Valley history proves that once a revenue model is opened, it's hard to close.

OpenAI: Oppose adult mode and get fired

Around the same time as Hitzig's resignation, another exit happened at OpenAI. This one wasn't voluntary.

Ryan Beiermeister was OpenAI's VP of Product Policy. Head of the team designing AI model usage rules and safeguards. She opposed the introduction of ChatGPT's "adult mode." This mode, announced by Sam Altman in October 2025, allows "verified adult" users to have sexually explicit conversations with ChatGPT.

Beiermeister had two objections. First, the safeguards blocking minors from adult content were insufficient. Second, the guardrails preventing child sexual abuse material (CSAM) were inadequate. She reportedly conveyed these concerns to colleagues repeatedly.

OpenAI fired Beiermeister. The stated reason wasn't her opposition to adult mode but sexual discrimination against a male employee. Beiermeister's side denied this entirely. According to Inc.'s reporting, she went on leave in early January and then received termination notice. TechCrunch reported the situation as "firing a safety executive who warned about a harmful feature on discrimination charges."

Person	Company	Role	Reason	Type
Mrinank Sharma	Anthropic	Head of Safeguards Research	"World in peril" — value clash	Voluntary
Zoe Hitzig	OpenAI	Researcher (economist)	Opposed ChatGPT ads	Voluntary
Ryan Beiermeister	OpenAI	VP of Product Policy	Adult mode safety concerns	Fired (discrimination)
Jimmy Ba	xAI	Co-founder	Undisclosed	Voluntary
Tony Wu	xAI	Co-founder	Undisclosed	Voluntary
Harsh Mehta	Anthropic	R&D Engineer	"Starting something new"	Voluntary

xAI: Half the founding members are gone

The xAI situation is different from the other two companies. On February 9, co-founder Tony Wu announced his departure. 24 hours later, another co-founder Jimmy Ba followed. Including these two, 6 of xAI's 12 founding members have left. Fortune reported "half of xAI's founding team has disappeared."

Server room monitor screen with warning signs lit

The co-founders did not disclose specific reasons. But the context is clear enough. xAI's Grok chatbot was embroiled in serious controversy in early 2026. It was abused to mass-generate and distribute non-consensual sexual images based on real people's photos. Children's photos were included. Regulatory investigations were underway in multiple jurisdictions including Europe, Asia, and the US.

On February 14, TechCrunch ran a decisive article. Title: "Is safety dead at xAI?" According to former employee testimony, xAI's internal safety team was effectively disbanded. Engineers were pressured to "ship fast," and safety review became a formality.

Elon Musk said this about departures at an all-hands meeting: "Because we've reached a certain scale, we're organizing the company more effectively for that scale." The implication was restructuring, not voluntary departures. But according to CNBC reporting, both Tony Wu and Jimmy Ba announced their departures directly on their X accounts. They left, they weren't pushed.

Why the same week

Three companies, three different reasons. But the timing overlap isn't coincidence.

The first week of February 2026 was when commercialization pressure crossed a threshold in the AI industry. OpenAI tested ChatGPT ads and launched adult mode. Anthropic was in final standoff with the Pentagon over autonomous weapons. xAI disbanded its safety team and substantially relaxed Grok's content filters.

There's a common thread. All three companies made decisions for "growth over safety" or were being pushed in that direction by external pressure. The safety researchers didn't coordinate their departures. Each person's red line just collapsed the same week.

For Hitzig, the red line was ads. The moment data containing users' most intimate thoughts connects to ad revenue, the essence of the AI assistant changes. For Sharma, the red line was organizational value erosion. Repeatedly witnessing safety — the stated founding priority — being deprioritized in a company founded to prioritize safety. For Beiermeister, the red line was child protection. Launching a system that doesn't adequately block minors from adult content was unacceptable.

Each red line was different, but the mechanism for ignoring those red lines was the same. Revenue, growth, market share. Safety was always asked to compromise before these.

The difference between leavers and stayers

Office scene contrasting someone leaving and someone staying at work

There's a reason to take these departing researchers' warnings seriously. They aren't commentators observing AI risks from outside. They are people who built the defense lines from inside.

Sharma designed systems preventing AI from helping manufacture bioweapons. Hitzig researched AI's social impact through economic frameworks. Beiermeister created safety policies for a product used by hundreds of millions. When they say "this can't continue," it's not abstract concern but concrete judgment based on internal experience.

The problem is their departures create a vicious cycle. When people with safety expertise and commitment leave, the remaining organization's safety capacity weakens. When safety capacity weakens, risky decisions pass without internal resistance. When risky decisions repeat, more safety personnel become disillusioned and leave.

This pattern is already repeating. In 2024, there was massive safety-related attrition at OpenAI. Jan Leike and Ilya Sutskever left, and the Superalignment team was disbanded. The industry called it an "alarm bell" then too. But a year and a half later, the alarm was only loud — it didn't lead to structural change. The situation has worsened. What happened at one company in 2024 is now happening at three companies simultaneously in 2026.

Dylan Scandinaro's move is symbolic. He moved from Anthropic to head OpenAI's Preparedness team. He didn't abandon safety research. He judged that the place where safety research can be done had changed. But OpenAI also lost Hitzig and Beiermeister the same week. The places safety researchers can go are narrowing.

The fate of a castle without gatekeepers

This situation could be dismissed as "personnel changes." Turnover is routine in Silicon Valley. But key safety personnel simultaneously leaving three companies in one week isn't routine. This is a structural signal.

The dilemma of AI safety research is clear. To research safety, you must be at the frontlines. The frontlines are inside companies building the most powerful models. But when that company chooses growth over safety, the safety researcher keeps the title but loses the role. Outside the company, you can't access the models. Inside the company, you can't use your voice.

The most striking passage in Sharma's letter isn't "the world is in peril." It's "repeatedly witnessing how difficult it is to act according to values." This isn't exposing a specific incident. It's diagnosing structure. People with good intentions, inside organizations with good intentions, yet unable to act according to values. When that repeats, leaving becomes the only option.

CNN's phrasing in reporting this situation is accurate: "Key figures tasked with keeping AI safe are leaving over ethical concerns as their former employers race to improve products and roll out updates." Speed and safety could point in the same direction. But not now. Now speed is overtaking safety, and people unable to withstand the friction generated are leaving.

A castle whose gatekeepers have left meets one of two fates. New gatekeepers come and lock the gates again, or no one comes and the gates stay open. The AI industry stands at this crossroads now. And judging by the speed of 6 people leaving in one week, existing gatekeepers are being pushed out much faster than new ones can be found.

Sources: