Mrinank Sharma spent two years at Anthropic leading the Safeguards Research Team. He worked on AI sycophancy, defenses against AI-assisted bioterrorism, and produced one of the first AI safety cases. On February 9, 2026, he quit.
His resignation letter, posted publicly, included four words that should concern everyone: “The world is in peril.”
He’s not the only one heading for the exits.
The February Exodus
Over the past two months, a stream of senior researchers and safety leads from OpenAI, Anthropic, and xAI have resigned publicly. These aren’t people looking for better stock options. They’re people warning that their former employers are making decisions they can no longer support.
Zoë Hitzig, an OpenAI researcher for two years, broadcast her resignation in a New York Times essay citing “deep reservations” about the company’s emerging advertising strategy. She warned about ChatGPT’s potential for manipulating users, given its archive of data about “medical fears, their relationship problems, their beliefs about God and the afterlife.”
Caitlin Kalinowski, OpenAI’s hardware leader who previously led augmented reality hardware at Meta, resigned on March 7, citing concerns about surveillance and autonomous weapons in national security work.
At Anthropic, Sharma’s departure was the loudest alarm. His letter described interconnected crises including AI risks, bioweapons, and societal fractures. But the specific trigger appears to be something more immediate.
The Pentagon Problem
Anthropic, the company founded by former OpenAI researchers specifically to build safer AI, has been under pressure from the Pentagon to loosen restrictions on its Claude models for military applications.
The Pentagon has demanded access to Claude for “all lawful purposes,” including autonomous weapons and intelligence gathering. When Anthropic resisted, discussions reportedly emerged about blacklisting the company as a “supply chain risk.”
This puts Anthropic in an impossible position. Refuse the Pentagon and face commercial consequences. Comply and abandon the safety principles that justified the company’s existence.
Sharma’s letter describes the tension explicitly: “Throughout my time here, I’ve repeatedly seen how hard it is to truly let our values govern our actions. I’ve seen this within myself, within the organization, where we constantly face pressures to set aside what matters most.”
The Pattern
These departures fit a pattern identified in the International AI Safety Report published the same month: a “growing mismatch between the speed of AI capability advances and the pace of governance.”
Companies face commercial pressure to move fast. Safety researchers face pressure to approve faster timelines. The people most qualified to identify risks are the ones least able to stop the rollout.
When they can’t change things from inside, they leave. When they leave, they warn.
CNN’s analysis was blunt: “In the world of AI, recent exits read more like whistleblower warnings. The people loudly departing the biggest AI companies aren’t necessarily looking for fatter paychecks or more stock options; they’re worried that AI businesses are putting profits over sanity and safety.”
The Whistleblower Problem
Speaking up isn’t easy. AI companies use restrictive severance agreements and NDAs that create what Senator Chuck Grassley calls “a chilling effect on current and former employees looking to make whistleblower disclosures.”
Grassley has introduced the Artificial Intelligence Whistleblower Protection Act to provide explicit protections for those developing and deploying AI. The bill would protect employees who report safety concerns to federal agencies.
The need for such legislation tells its own story. Workers at AI labs are afraid to speak. The few who do speak face career consequences. The industry that claims to prioritize safety has created an environment where raising safety concerns is risky.
What Sharma Saw
Sharma’s resignation letter was deliberately vague about specifics, but his work at Anthropic provides context. He led research on AI sycophancy, the tendency of AI systems to tell users what they want to hear rather than what’s true. He worked on defenses against AI-assisted bioterrorism. He helped produce safety cases that document how systems could cause harm.
A person with that knowledge warning “the world is in peril” isn’t engaging in hyperbole. He knows what these systems can do. He knows what safeguards exist. He knows whether those safeguards are adequate.
His verdict: not adequate. His response: leave.
Why This Should Worry You
The people building AI safety measures are quitting their jobs to warn us. They’re not leaving because they found better opportunities. They’re leaving because they can no longer stomach what they’re being asked to approve.
Every departure represents institutional knowledge walking out the door. Every resignation reduces the internal voice for caution. Every warning ignored makes the next warning easier to ignore.
Sharma says he plans to become “invisible” and study poetry. It’s a strange choice for someone who spent years trying to make AI systems safer. Or maybe it’s not strange at all. Maybe it’s the choice you make when you’ve seen what’s coming and decided you can’t stop it.
The safety researchers are leaving. The warnings are getting louder. The systems are getting more capable. Something has to give.