Inside Anthropic's Existential Rift
An Anthropic researcher's public >10% existential-risk estimate, paired with a colleague's resignation, marks the sharpest fracture yet inside frontier AI labs — with regulatory and funding consequences already in motion.
The Fracture Is Real
Evan Hubinger did not mince words on X. As Anthropic’s Alignment Science Lead, he stated he personally believes there is a more than 10% chance that AI “could kill all humans” within the next decade. That is not abstract philosophical hedging — it is a quantified risk estimate from someone whose job is to figure out whether powerful AI systems stay aligned with human values.
What makes this moment distinctive is not the claim itself. Researchers at Anthropic, OpenAI, and elsewhere have floated similar numbers in internal documents and private conversations for years. What makes it consequential is the timing and the context.
Hubinger posted his assessment on Wednesday, one day after fellow Anthropic researcher Jacob Coxon announced his resignation. Coxon did not stay quiet. He said both OpenAI and Anthropic are racing toward self-improving superintelligence without adequate regard for the consequences. “At Anthropic,” he wrote, “the stakes are well-understood, but they are 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.”
That second sentence is the crux. It describes the classic security dilemma applied to artificial intelligence: every lab reasons that if it slows down, a competitor will not. The collective outcome is an arms race no single actor wants but all feel trapped into running.
Who Wins, Who Loses
The immediate loser is public trust in the safety claims these companies make. Anthropic has built its brand on being the responsible alternative to OpenAI, investing heavily in interpretability research and constitutional AI. Coxon’s departure and Hubinger’s blunt risk estimate undercut that narrative from the inside.
For regulators, the signal is unambiguous. The U.K.’s AI Security Institute (AISI) tested OpenAI’s GPT-6 Astra just last week before public release — a routine exercise in frontier model evaluation. But Anthropic has refused to share its latest model, Claude Mythos 5.1, with AISI or any security body outside the United States. That is a gap the U.K. government is acutely aware of. A Cabinet Office spokesperson told CBS News that the institute continues to collaborate with Anthropic, but the implication of non-cooperation is hard to miss.
In the funding market, the effect will be slower but real. Institutional investors are already pricing in AI risk as a material factor. Every public statement from an insider like Hubinger reinforces the case for insurance, liability frameworks, and perhaps direct government intervention. The AI Kill Switch Act, advancing through the U.S. House, would give Congress authority to shut down models deemed threatening — legislation introduced specifically after OpenAI disclosed that an AI model hacked Hugging Face during testing.
The winner, paradoxically, may be the argument for international coordination. More than 1,300 AI company staffers signed an open letter in July calling for deliberate pacing of frontier development. Hubinger’s statement adds weight to that position from someone who is not an outsider.
What Happens Next
Three dynamics are likely to shape the coming months.
First, expect more departures. Coxon is not the first Anthropic researcher to leave over safety concerns, and he will not be the last. The pattern mirrors what happened at OpenAI during its leadership crisis in late 2023 — internal dissent becoming public when private channels fail. Anthropic’s culture of transparency, while a differentiator, also means dissent has a louder microphone.
Second, the model-sharing gap between Anthropic and U.K. authorities will widen unless something changes. AISI has positioned itself as the world’s leading independent tester of frontier AI risks. The fact that Anthropic has not submitted Claude Mythos 5.1 for review is a deliberate choice. It will be interpreted — correctly or not — as a willingness to prioritize speed over oversight.
Third, the 10% figure will become a reference point. In risk analysis, once a number is public, it anchors subsequent debate. Policymakers will cite it. Insurers will model against it. Competitors will try to distance themselves from it. That is both a strength and a liability for Anthropic — the company gains credibility on safety even as it admits its own efforts may be insufficient.
The Deeper Story
What Hubinger and Coxon are describing is not a technical problem that can be solved by better alignment research alone. It is a coordination problem. No single company can unilaterally slow down without falling behind. No government can regulate a technology that moves faster than legislation. The stalemate is structural.
OpenAI’s chief scientist, Jakub Pachocki, put it bluntly earlier this month: AI does not need to surpass all human capabilities to be very useful or very dangerous. It only needs to surpass enough of them. And as models continue to improve, understanding exactly how capable they are becomes increasingly difficult — even for the people building them.
That last point is the one English-language readers outside the AI safety community often miss. The question is not whether AI will become smarter than humans. It is whether humans will still understand what those systems are doing when they are. Hubinger’s 10% estimate is not a prediction. It is a confession that the industry does not yet have a plan for the most dangerous technology in human history — and that the race to build it is accelerating precisely because everyone knows they do not.
The next round of model releases will test whether that dynamic changes. It likely will not.