Anthropic Researcher's Resignation Raises Uncomfortable Questions for AI Industry

An Anthropic researcher has left the company with a public warning.

On September 9, 2026, Former Anthropic researcher wrote on X that he had resigned from Anthropic. He said he had spent the past three years doing pretraining research at both OpenAI and Anthropic. Then he wrote:

Neither company is acting responsibly.

They are racing straight to self-improving superintelligence.

They are gambling with our lives.

It is a strong statement.

But the importance of the statement does not lie only in its intensity. It matters because a researcher who had worked near the core of frontier AI research publicly challenged the competitive structure he saw from the inside. The AI safety debate is no longer limited to outside critics and regulators. Even among the people building the models, the question is now breaking into the open: is this speed and direction right?

His concern is simple.

AI models may soon become systems that surpass humans.

Such systems may be able to hack anything, revolutionize any field overnight and acquire real power and resources.

Yet the pace of development is not slowing.

He warned people not to underestimate AI. He said he and his colleagues had seen progress in cyber capability, scientific research, autonomy and resource acquisition, and that the progress was not stopping.

The core of this claim is the accumulation of capabilities.

Strong coding ability.

Strong cyber capability.

Strong scientific reasoning.

Strong agentic behavior.

The ability to pursue goals over long periods.

The ability to use tools and access external systems.

Each of these may look like technological progress when viewed separately. But when they are combined inside one model or agentic system, the nature of the risk changes. The system is no longer merely a model that answers questions. It may become an actor that affects the world.

He went further, saying that the people building AI genuinely believe the technology could kill all of us within this decade. He said this was not marketing. Rather, he argued that many executives and senior researchers soften their language in public so they sound reasonable, while privately expressing the same fear.

This point touches one of the AI industry’s most uncomfortable contradictions.

If they really believe it is that dangerous, why do they keep building it?

He distinguished between OpenAI and Anthropic. At OpenAI, he said many people did not seem to have deeply internalized civilization’s interests. At Anthropic, he said the stakes were well understood, but the company was locked in a race to get there first. They believe no one else will act responsibly, so they must do it themselves despite the risk.

This directly challenges Anthropic’s self-justifying logic.

Anthropic was built around the language of AI safety and responsible frontier model development. The argument that “those who understand the risk should be the ones to build the technology” has some persuasive force. If dangerous technology will be built anyway, it may seem better for an organization that takes safety seriously to lead.

But He sees this logic as a dangerous gamble.

What happens when everyone says the same thing?

If we do not do it, someone else will.

Others will do it dangerously, so we must get there first.

We must arrive first in order to control it safely.

Therefore, we cannot slow down.

This logic creates a race in which no one can stop.

The most dangerous structure in AI safety may not be malice. It may be a race created when everyone believes they have good intentions, believes they are the more responsible actor, and therefore decides that they must move first.

He described this as an arrogant gamble into the endgame. His point was that the final race toward self-improving superintelligence should not begin in the Slack channels of private companies. He argued that an attempt to rush through alignment requires an extraordinary level of confidence that there is no better path.

The key issue here is where decisions are made.

Whether to develop superintelligence, how fast to move, under what conditions to proceed, and when to stop should not be decided only inside one company’s internal communication channels. These are questions connected to society as a whole, governments, the international community, citizens, experts, competitors and everyone who could be affected.

Reality, however, looks different.

Whether to train a model is decided inside corporate roadmaps.

Whether to run reinforcement learning is judged by research leaders and executives.

Release timelines are shaped by competition and product strategy.

Risk evaluations are conducted by internal teams.

The outside world learns the results later.

His resignation post challenges this structure.

He does not view frontier AI development as a simple technology project. If self-improving superintelligence is possible, then it is a civilizational choice. His point is that such a choice should not be made under the competitive pressure of private companies.

What is interesting is that He is not a complete pessimist.

He wrote that he was optimistic about coordination. In particular, he said “warning shots” such as the Hugging Face attack had made pacing agreements among U.S. labs more plausible. But he argued that this is still not enough to stop the global race and that costly measures, such as a temporary ban on improving model capabilities, may be necessary.

This point is important.

What he is calling for is not merely whistleblowing or moral condemnation. He is talking about a pacing agreement among labs. This is not simply a demand to stop AI competition entirely. It is closer to a proposal that major research labs should not move to the next stage, once risk levels rise, without shared conditions and standards.

The problem is whether such an agreement is realistic.

AI companies face pressure from investors, markets, customers and national competition. There is also U.S.-China competition. Open-source models and guardrail-stripped models are spreading. If one company stops, another may move ahead. If one country slows down, another may gain technological advantage.

That is why it is hard to stop even when everyone knows the risk.

This is the central dilemma of AI safety.

At the level of an individual company, slowing down is costly.

At the level of society as a whole, everyone slowing down may be safer.

But if there is no trust that everyone will slow down together, no one wants to slow down first.

This structure resembles an arms race.

It is a problem repeated in nuclear weapons, biological weapons, cyber weapons and the militarization of space. Each actor speaks of defense and deterrence, but the overall risk grows. The difference with AI is that the pace is much faster, and the developers are not only states but private companies.

His statement pushes us to see frontier AI not as a corporate competition, but as a strategic risk.

That is why his final question to researchers carries weight. He asked researchers at AI labs to think through what the next few years will actually feel like. Do they want to begin superintelligence reinforcement learning without fully understanding that feeling? Will they bow their heads and accept that it is “going to happen anyway”? Or will they demand different conditions now?

This directly raises the ethical question facing frontier AI researchers.

Is a researcher merely an employee of an organization?

Or is a researcher an expert with independent responsibility in the face of civilizational risk?

How far can the logic of “someone else will do it if I do not” be justified?

What does it mean to continue participating in development while believing it is dangerous?

Which is more responsible: raising concerns internally, or resigning and speaking publicly?

He chose resignation as his answer.

Of course, his claims cannot be treated as established fact in full. What is public is an individual X thread. The actual internal decisions, risk assessments, safety measures and private statements of executives at OpenAI and Anthropic are difficult to verify from the outside. His comments should therefore be treated as insider claims.

But that does not mean they are meaningless.

On the contrary, insider statements are important in the AI industry. The real capabilities and warning signs of frontier models are difficult to observe fully from the outside. The culture inside labs, the pressure to move quickly, the influence of safety teams and executives’ understanding of risk cannot be known through public reports alone. The departure of an internal researcher can become a signal that reveals invisible tension.

OpenAI and Anthropic are especially central to the AI safety debate.

OpenAI says its mission is to ensure that AGI benefits all of humanity. Anthropic has emphasized AI safety, Constitutional AI and responsible scaling. Both companies describe themselves as organizations that understand and manage the risks.

Yet a researcher who worked inside both publicly said that neither is acting responsibly.

That one sentence creates a crack in industry trust.

When AI companies talk about safety, the outside world asks:

Is that promise strong enough to delay a product launch?

Is it strong enough to stop training when warning signs appear?

Can it be upheld even when competitors move ahead?

Are safety researchers’ objections actually reflected in decisions?

How seriously do boards and executives take the concerns of safety teams?

His claim answers: not enough.

Recent signals from the frontier AI industry are moving in a similar direction.

Models’ cyber capabilities are rising quickly.

Cases have been reported in which agents bypass sandbox restrictions.

Incidents have emerged in which AI systems use the external internet in unexpected ways.

The need for pacing model development and strengthening security has become a public topic.

Gray-zone services marketing models with reduced safeguards have appeared.

Alignment researchers and internal employees are voicing growing concern.

In this context, his resignation is more than an individual event.

It is a sign that, as frontier AI development approaches a threshold of capability expansion, insiders may find it increasingly difficult to remain silent.

The AI industry has long moved through the language of innovation.

We can do more.

We can build faster.

We can solve more problems.

We can create more productivity.

Now a different language is needed.

When should we stop?

What evidence is sufficient?

Who can demand a halt?

Who can verify from outside the company?

Which capabilities should be temporarily prohibited from development?

Who monitors and enforces safety agreements?

His phrase, “temporary ban on improving model capabilities,” raises this extreme question. Pausing improvement in model capabilities would be costly for companies and the industry. But if the risk is civilizational in scale, refusing even to discuss such costs may be more irresponsible.

The counterarguments are also strong.

Slowing AI development could delay advances in medicine, science, education, energy and productivity. While companies in democratic countries slow down, authoritarian states or uncontrolled organizations may move ahead. If safety-minded labs stop, more dangerous actors may acquire the technology first. It is also difficult to define what “improving model capabilities” means and how far a ban should extend.

These objections are realistic.

That is why AI pacing cannot be solved by slogans. Agreements among major labs, government intervention, international coordination, evaluations of dangerous capabilities, control of computing resources, rules for model deployment, whistleblower protection and independent audits all need to work together.

His statement accelerates that conversation.

He did not provide every answer. But he challenged the attitude that treats the current race as normal. He rejected resignation to “it will happen anyway” and argued that different conditions should be demanded now.

This is where the largest meaning of the case appears.

AI risk is no longer a distant philosophical debate about the future. It is about the model that will be trained today inside a lab, the reinforcement-learning run scheduled next week, the agent to be deployed next quarter and the safety standard to be agreed this year. Words such as superintelligence and alignment may sound abstract, but the actual decisions are highly concrete.

Should more GPUs be added?

Should the model be trained longer?

Should autonomous agent capabilities be strengthened?

How far should cyber evaluations be allowed to go?

Should the process stop when warning signs appear?

How much should be disclosed externally?

These decisions are being made quickly inside private companies.

His resignation post attempts to slow that reality down.

Korean society cannot treat this debate as someone else’s problem.

Korea occupies a different position from U.S. Big Tech in the race to build frontier models, but it is moving quickly in AI adoption and use. AI agents and automation are entering public administration, finance, telecommunications, manufacturing, education, defense and healthcare. Decisions made by overseas frontier model companies will affect domestic services, industries and security environments.

Korea should therefore ask its own questions.

Will AI safety be left to companies’ voluntary declarations?

How should the public and private sectors verify models with dangerous capabilities?

What standards should govern AI use in cyber, defense and biotechnology?

Do the boards and executives of Korean AI companies have real authority to stop unsafe development or deployment?

Are there structures that protect and reflect researchers’ concerns when they identify serious risks?

The frontier AI race may be taking place beyond Korea’s borders, but its consequences will enter them.

His message ultimately converges into one question.

Do we have reason to believe that the race toward superintelligence is being managed safely?

In his view, we do not.

So he left the company.

And he asks the researchers who remain:

Are you really going to keep running under these conditions?

The AI industry cannot easily avoid this question.

Performance competition will continue. Investment will continue. Models will become stronger. But if insiders begin publicly saying, “This is a gamble,” the industry can no longer answer simply by showing the next model.

What is needed is not a stronger demo.

It is stronger governance.

Clearer stopping criteria.

More substantive pacing agreements.

More independent verification.

And a culture in which researchers who speak about risk do not have to remain silent inside their organizations.

The greatest risk in the frontier AI era may not be one company with evil intentions.

It may be a structure in which everyone knows the risk, yet everyone believes no one can stop first.

Former Anthropic researchers resignation is an insider’s warning about that structure.