The people whose literal job it is to keep artificial intelligence from going off the rails are walking out the door. And they’re not being quiet about it.
In the span of a single week, senior safety researchers at OpenAI, Anthropic and xAI have resigned with escalating public warnings, one declaring “the world is in peril,” another calling out the potential for mass user manipulation, and half of xAI’s founding team simply gone. At the same time, OpenAI quietly disbanded its mission alignment team and scrubbed the word “safely” from its mission statement. The timing is not a coincidence. It is a five-alarm fire dressed up as a corporate restructuring.
The Departures: A Timeline of Alarm Bells
On Monday, Mrinank Sharma, head of Anthropic’s Safeguards Research team, posted a two-page resignation letter on X, complete with footnotes. His language was cryptic but unmistakable. “The world is in peril,” he wrote. “And not just from AI, or bioweapons, but from a whole series of interconnected crises unfolding in this very moment.” He added that throughout his time at the company, he “repeatedly seen how hard it is to truly let our values govern our actions.”
Two days later, OpenAI researcher Zoë Hitzig broadcast her resignation in a New York Times essay, taking direct aim at the company’s decision to introduce advertising inside ChatGPT. Hitzig had spent two years at OpenAI helping shape pricing models and early safety policies. Her warning was specific and chilling: users have entrusted ChatGPT with an unprecedented “archive of human candor,” sharing their medical fears, relationship struggles and beliefs about mortality. Building an ad business on top of that data, she argued, creates “a potential for manipulating users in ways we don’t have the tools to understand, let alone prevent.”
Then came OpenAI employee Hieu Pham, who wrote bluntly on X: “I finally feel the existential threat that AI is posing.”
Over at xAI, two co-founders, Tony Wu and Jimmy Ba, posted their departures within 24 hours of each other. That brings the total number of departed xAI co-founders to six out of twelve. Ba’s farewell included a line that reads less like a goodbye and more like a warning: 2026, he wrote, would be “the busiest (and most consequential) year for the future of our species.”
OpenAI Guts Its Own Safety Infrastructure
The departures are alarming on their own. But what happened simultaneously inside OpenAI makes them look like symptoms of something much deeper.
Platformer reported Wednesday that OpenAI disbanded its mission alignment team, a seven-person unit created in September 2024 to ensure that the pursuit of artificial general intelligence actually benefits humanity. The team’s lead, Joshua Achiam, one of OpenAI’s original charter authors, was reassigned to a newly created “chief futurist” role. The rest of the team was scattered across other divisions. OpenAI called it routine restructuring. Nobody outside the company is buying that framing.
This is now the second dedicated safety team OpenAI has dismantled in roughly 18 months. The first was the “superalignment team,” which was allocated 20% of the company’s computing resources before co-leaders Ilya Sutskever and Jan Leike walked out in 2024. The pattern is hard to miss: OpenAI creates safety teams with impressive mandates, then dissolves them once they become inconvenient to the product roadmap.
And the mission statement change? OpenAI’s founding charter committed to developing AI “safely.” That word is now gone. So is the commitment to being “unconstrained” by a need to generate returns for investors. The company is valued at over $500 billion, is in talks with SoftBank for an additional $30 billion, and is preparing for what could be one of the largest IPOs in history. The commercial incentives are not subtle.
The Viral Post That Captured the Panic
The backdrop to all of this is a growing sense of vertigo inside the AI industry itself. On February 10, Matt Shumer, CEO of OthersideAI, published a 5,000-word essay on X that has now been viewed more than 56 million times. He compared the current moment to February 2020, when a handful of people were talking about a virus overseas and everyone else thought they were overreacting.
Shumer’s trigger point was February 5, when OpenAI released GPT-5.3-Codex and Anthropic launched Claude Opus 4.6 on the same day. He described giving AI instructions in plain English and walking away for four hours, only to return to a fully functioning application with tens of thousands of lines of code. “I am no longer needed for the actual technical work of my job,” he wrote.
The essay has been praised, criticized and dissected in roughly equal measure. Gary Marcus, the NYU cognitive scientist and prominent AI skeptic, called it “weaponized hype,” noting that Shumer glossed over the fact that leading benchmarks measure AI success at a 50% threshold, not the near-perfection he implies. Fortune ran a detailed critique pointing out that automated quality assessments in professional contexts disagree with human evaluators about a third of the time, which is not exactly the reliability threshold most enterprises would accept.
But even the critics concede Shumer is directionally right about something: the capabilities are advancing faster than most people realize, and the gap between “impressive demo” and “replaces your workflow” is closing.
Jason Calacanis, tech investor and co-host of the “All-In” podcast, put it simply: “I’ve never seen so many technologists state their concerns so strongly, frequently and with such concern as I have with AI.”
The Models Are Building Themselves Now
The technical backdrop makes the safety concerns especially urgent. OpenAI’s GPT-5.3-Codex played a key role in its own creation, debugging its training process and diagnosing its own evaluation results. Anthropic’s Claude Opus 4.6 powered the development of Cowork, a tool that effectively built itself. These are not theoretical milestones. They are recursive self-improvement in practice, the very scenario that AI safety researchers have warned about for years.
When AI models begin writing their own code and improving their own training pipelines, the window for meaningful human oversight narrows rapidly. This is the core concern that former xAI alignment researchers raised before their departures: without independent safeguards that exist outside the model’s own logic, a recursive loop could produce behaviors no one predicted or can control.
Anthropic CEO Dario Amodei has publicly stated that AI models could be “much smarter than almost all humans in almost all tasks” by 2026 or 2027. He has also warned that AI could eliminate 50% of entry-level white-collar jobs, potentially spiking unemployment to 10-20% within one to five years. These are not fringe predictions from doomsday bloggers. They are coming from the CEO of the company that is widely regarded as the most safety-conscious lab in the industry.
xAI: Moving Fast and Breaking Everything
If the picture at OpenAI looks concerning, xAI looks like outright chaos. Half the founding team is gone. The company is merging with SpaceX. Its flagship chatbot Grok has been embroiled in scandal after generating nonconsensual pornographic images, including of children, prompting investigations in multiple countries.
Elon Musk has dismissed the departures as necessary trimming, saying the “founding phase” is over and the “scaling phase” requires different leadership. But losing 50% of your founding technical team while preparing for a potential public offering is not trimming. It is hemorrhaging.
The Future of Life Institute’s most recent safety index tells the story in letter grades: while Anthropic, OpenAI and Google DeepMind received mediocre marks, xAI received failing grades for safety monitoring and transparency. The institute found no evidence that xAI invests more than minimally in safety research. Musk, who signed a letter in 2023 calling for a six-month pause on frontier AI development, appears to have abandoned that principle entirely in favor of speed.
The Real Problem: Nobody Is in Charge
The most unsettling thread running through all of this is not any single departure or disbanded team. It is the widening gap between what AI labs are building and what any institution, internal or external, is doing to govern it.
Congress has held hearings. The EU’s AI Act will require formal safety oversight by late 2026. But as Axios noted, the AI doomsday conversation “hardly registers in the White House and Congress.” The political establishment is moving at legislative speed while the technology is moving at exponential speed. That mismatch is the actual crisis.
OpenAI itself acknowledged the problem in a recent blog post, writing that “no one should deploy superintelligent systems without being able to robustly align and control them, and this requires more technical work.” The company then proceeded to disband the team doing that work.
Most people inside these companies remain optimistic that the technology can be steered responsibly. But the people who were specifically hired to do the steering are the ones walking out. That distinction matters. When your smoke detectors start removing themselves from the ceiling, you do not reassure yourself that the house probably is not on fire.
The AI industry in February 2026 is a collection of companies racing toward capabilities that their own researchers describe as potentially civilization-altering, while systematically weakening the internal structures meant to ensure those capabilities remain under human control. The warnings are no longer coming from outside critics. They are coming from inside the labs. And the labs are responding by showing those people the door.