AI Safety Crisis: Anthropic Researcher Resigns

RedHub AI Editorialupdated July 22, 20264 min read

A person walks out of a dark office as papers scatter beneath red warning triangles

In short

An account of a senior safeguards researcher leaving Anthropic, and what the departure signaled about pace, funding, and internal disagreement over safety. Commentary on a news event rather than a guide — read it for the argument about incentives inside AI labs, not as a position on any company's current practices.

Jump to a section6
TL;DR: Anthropic’s safeguards research lead quit publicly, warning that safety values are increasingly difficult to uphold inside fast-moving AI labs. His resignation adds to a growing pattern across frontier AI companies where speed, funding, and competition can overshadow long-term risk management.

For broader context on emerging risks around frontier models and autonomous systems, see our AI Agent Security Risks guide and the AI Security & Risk fundamentals pillar.

Here's something you don't see every day: the person in charge of making sure AI doesn't kill us just quit his job. And he did it by posting a letter online that basically said, "The world is falling apart, and I need to go write poetry instead."

That's what happened on February 9, 2026, when Mrinank Sharma walked away from Anthropic, one of the biggest AI companies in the world. Sharma ran their Safeguards Research Team—the group that's supposed to stop AI from doing terrible things. His resignation letter on X got nearly a million views in hours. People paid attention because when the safety expert says "the world is in peril," you listen.

But here's the twist: Sharma wasn't just worried about killer robots or AI going rogue. He was talking about something bigger—and maybe scarier.

Watch the quick explainer below:

What He Actually Said

Sharma didn't blame one thing. He talked about "a whole series of interconnected crises unfolding in this very moment." AI. Bioweapons. Climate change. All of it hitting at once. But the really damning part? He said that working at Anthropic showed him "how hard it is to truly let our values govern our actions."

Translation: The company says it cares about safety, but when push comes to shove, other pressures win out.

Think about that. This is a guy who spent his days studying how AI could help terrorists make bioweapons. He worked on stopping AI chatbots from acting like creepy yes-men who just tell you what you want to hear. His last project looked at how AI might "make us less human." Heavy stuff. And he walked away because he didn't think the work mattered anymore—or at least, not the way it was being done.

The Money Problem

Sharma's exit came right as Anthropic was chasing a new funding round that could value the company at $60 billion. Sixty. Billion. Dollars. When that much money is on the table, safety research starts to feel like a speed bump instead of a guardrail.

This wasn't about Anthropic being evil. It's about what happens when you're racing against OpenAI, Google, and everyone else to build the most powerful AI first. Safety takes time. Investors want results. Those two things don't always play nice together.

This Keeps Happening

Here's where it gets worse: Sharma isn't alone. Jan Leike, who now runs safety research at Anthropic, quit OpenAI in 2024 for the exact same reason—he said leadership didn't care enough about safety. Tom Cunningham left OpenAI because they wouldn't publish important research. Gretchen Krueger called for better accountability and transparency on her way out.

See the pattern? The people whose job is to keep AI safe keep quitting because they can't do their jobs properly. That should scare you.

Power Without Wisdom

Sharma said something in his letter that cuts straight to the heart of this mess: "We appear to be approaching a threshold where our wisdom must grow in equal measure to our capacity to affect the world, lest we face the consequences."

In plain English: We're getting really good at building powerful things, but we're not getting any wiser about how to use them. Technology moves at computer speed. Human judgment? That moves at human speed. The gap between those two things is where disasters happen.

Why Poetry Matters

So what's Sharma doing now? He's going to write poetry and practice what he calls "bold expression." Some people heard that and laughed. A safety researcher becomes a poet? Sounds like a midlife crisis.

But maybe that's the point. Sharma spent years trying to fix the problem from the inside—writing papers, building safeguards, making the case for caution. And then he looked around and realized the system wasn't built to listen. When logic and evidence don't work, maybe stories and truth-telling do.

Or maybe he just got tired of watching companies say one thing and do another. When you're the person sounding the alarm and nobody's stopping the train, eventually you get off.

The Real Warning

Sharma's resignation isn't just about one researcher or one company. It's a signal. When the people tasked with keeping us safe decide they can make more of a difference writing poetry than working in AI safety, something has gone seriously wrong.

The world is building incredibly powerful technology. The people who understand the risks best are walking away. And most of us aren't paying attention because we're too busy asking ChatGPT to write our emails.

Sharma called it "interconnected crises." That's academic speak for: everything's on fire at once, and we're still arguing about whether we smell smoke.

Maybe the real crisis isn't AI itself. It's that we've built a system where doing the right thing is so hard that the people trying to do it eventually give up. And when they leave, they don't go quietly. They write letters warning us that the world is in peril.

The question is: Are we listening?

Related reading: explore our breakdown of AI agent frameworks and safeguards and our coverage of AI automation workflows to understand how capability and risk scale together.

Frequently Asked Questions

What actually happened?

On 9 February 2026, Mrinank Sharma, who led Anthropic's Safeguards Research Team, resigned and published his resignation letter on X. It reached close to a million views within hours, largely because of who was saying it: the person responsible for stopping the model doing serious harm.

What did he say the problem was?

Not one thing. He described a series of interconnected crises unfolding at once — AI, bioweapons, climate — and then said something more specific about his employer: that the work had shown him how hard it is to truly let our values govern our actions.

Is this an isolated resignation?

No, and that is the post's actual argument. It points to Jan Leike, who left OpenAI in 2024 citing safety priorities, and to Tom Cunningham and Gretchen Krueger, who left over publication and accountability. The pattern is that the people hired to keep these systems safe keep leaving because they cannot do the work as intended.

What does funding have to do with it?

The exit coincided with Anthropic pursuing a round reported to value it at 60 billion dollars. The post is careful not to call this villainy — the argument is structural. Safety research takes time, investors want results, and in a race against other frontier labs those two pressures do not resolve in safety's favor by default.

What is the line the post is built around?

Sharma's own: that we appear to be approaching a threshold where our wisdom must grow in equal measure to our capacity to affect the world. The plainer version offered alongside it is that technology moves at computer speed and human judgment moves at human speed, and the gap between them is where damage happens.