The 10% Nobody Can Short: Anthropic’s Resignations and the Missing AI Risk Model

Jacob Coxon delivered the most consequential risk disclosure of the AI cycle, then walked away from it. On X, the Anthropic safety researcher put a number on a nightmare: artificial intelligence, he claims, has a greater-than-10% probability of "killing all humanity." Then he resigned. His exit came hours after another Anthropic employee quit with a phrase that reads like a margin call — the lab, she said, is "gambling with our lives." Coxon was blunter about the mechanism: Anthropic and OpenAI are "racing toward self-evolving superintelligence," staking our existence as the cost of entry. Ten percent is not a rounding error. In any properly risk-managed book, a double-digit probability of terminal ruin is a position-sizing catastrophe. It demands a model, assumptions, stress tests, and a published disclosure. None has appeared. No methodology. No simulation. No base rates. Just a number, a tweet, and two empty chairs at the safety desk. Anthropic exists because of a schism. Its founders walked away from OpenAI in 2021, convinced that capability was outrunning caution. They built Claude around Constitutional AI — an attempt to make "helpful, honest, harmless" an engineering property rather than a marketing slogan. In a crypto ecosystem rightly suspicious of centralized AI power, Anthropic was the closest thing to an ethical counterweight. Now add the compound implication. The safety-first lab is losing its safety people over safety concerns. When a researcher who spends his days probing these systems publicly puts the probability above 10% and then leaves, one of two things is true. Either he is wrong, or the rest of us have been making decisions about AI adoption, enterprise integration, and token exposure without information that should have been priced in long ago. The phrase "self-evolving superintelligence" deserves technical scrutiny. It assumes frontier models are approaching recursive self-improvement — systems that modify their own code and architecture without human gatekeeping. That is not an established fact; it is a hypothesis with extreme asymmetries. Yet the two best-funded labs in the world behave as though it is a near-term certainty, burning billions in compute for first-mover advantage. This is not a disagreement about ethics. It is a disagreement about timing — and in markets, timing disagreements are the only kind that matter. Evaluate the claim like a trade. Premise: existential catastrophe risk sits above 10%. The defensive entry would be slowing development, killing a training run, or building external oversight. The risk of acting: if the warning is wrong, another lab or another jurisdiction builds the capability first under weaker constraints. And the exit — the condition that closes the trade — is documented, verifiable evidence that the risk is being managed. The resignations are not that exit. They are exits without an underlying hedge. Worse, they leave the position open. Anthropic and OpenAI will hire other researchers, many of whom will conclude that being inside the room is safer than being locked out. The deeper problem is the number’s provenance. Ten percent is not a risk assessment; it is a claim waiting for a discipline that does not yet exist. Was it derived from internal red-team tests? Monte Carlo simulations of a recursive-improvement cascade? Expert elicitation among colleagues? A gut feeling developed after too many late nights reading about reward hacking? No one outside the room can tell. That absence of provenance is more dangerous than the number itself, because it converts a measurement problem into a faith proposition. In my own trading, I treat every unaudited probability as noise until it carries a position size. The financial equivalent would be a banker announcing a 10% default probability while producing no balance sheet, no cash-flow model, no covenant analysis. No credit committee would approve that loan. Yet in AI governance, the same epistemic standard is treated as whistleblowing. This is where my own history makes me allergic to unverified warnings. In 2017, I manually audited ERC-20 contracts for two mid-cap ICOs and found reentrancy vulnerabilities in the TokenSale contracts of projects that had raised over €5 million combined. I forked the code and demonstrated the exploit to founders directly. The exercise was not about dramatic predictions; it was about showing a failure mechanism in code that anyone could verify. That is the difference between an auditor and an alarmist: reproducible evidence. Terra’s code was poetry; Luna’s exit was prose. Do you remember the confidence? Do you remember how many qualified people insisted the UST mechanism was sound? The on-chain data told a different story before the collapse. Liquidity was leaving; pools were thinning; the supposed risk-free yield was pulling in exactly the counterparties who would flee first. I exited my stablecoin positions in May 2022 at the first sign of that drain and watched the cascade from the outside. That lesson now shapes every protocol I assess: evaluate a system by its exit, not its intention. Who can leave, when, and what happens to the ones who cannot? Apply the questions to the superintelligence race. Employees can exit, so they resign. Private investors can exit if they find secondary buyers. Institutions can exit by delaying integration. The rest — billions of people outside the lab — cannot exit at all. Any risk with no exit for the majority of its counterparties demands a higher standard of proof than a departing researcher’s anecdotal decimal. Where is the verification layer? In crypto, we built audits, open-source code, on-chain analytics, liquidation engines, and formal proofs. The AI industry offers ethics statements, model cards, and safety frameworks. A constitutional AI is an internal constraint. It is not independently audited and it is not tamper-proof. There is no block explorer for a neural network, no on-chain proof that a reward function contains no lethal loophole. Options don’t care about your conviction; they mark to market. If a model cannot be externally stress-tested, no internal whitepaper can guarantee its tail behavior. Then in 2026 I piloted an AI-trading system with a Paris startup and saw the problem up close. The model processed news sentiment faster than any human. It also hallucinated three trade executions that I had to cancel manually. If a model cannot be trusted to place a €500,000 options order without a human auditor, it cannot be trusted to rewrite its own optimizer. That gap between speed and judgment is the entire risk surface. So what does the 10% claim actually signal? Not a measured probability — a cultural one. It signals that people closest to the problem consider the risk real and unmanageable enough to justify resignation. But quitting a lab is not managing the risk. It is a trader leaving the building while the position runs against everyone else. Risk isn’t a number on a dashboard; it is the gap between belief and reality — and that gap is enormous. The contrarian reading is worth stating, because the easy narrative — safety people flee, therefore AI is dangerous — is itself a trap. First, selection bias runs through every safety researcher’s probability estimate. Someone who spends years studying AI catastrophe is not a random observer; he holds a strong prior that catastrophe deserves attention. That does not make his numbers wrong. It makes them, by construction, something other than a neutral posterior. Insiders are often the most contaminated witnesses to tail risk, not the most reliable ones. Second, the resignations may be net bullish for Anthropic and OpenAI in a perverse way. The loudest safety voices inside were friction. Without them, leadership faces less resistance as it pushes toward capability milestones. That is bullish for shipping products and bearish for everyone else. Third, "self-evolving" is doing too much undefined work. Current frontier systems remain constrained by human-designed architectures and supervised feedback loops. A model that uses a tool to edit code is not the same as a model that rewrites its own optimizer. Conflating the two inflates near-term catastrophe risk and drains urgency from the dangers that already exist: bias, autonomous cyber operations, disinformation, financial manipulation, and the concentration of power in a handful of unaccountable corporations. And a fourth observation, aimed directly at my own industry. We have spent years rejecting the idea that issuers can freeze our assets on a whim, and we have watched regulators sanction code itself — the Tornado Cash precedent treated open-source software as a criminal actor. If a culture that distrusts Circle’s ability to freeze USDC addresses accepts a 10% extinction claim without demanding its model, it is not the skeptic community it claims to be. The blockchain toolkit — transparency, auditability, immutability, decentralized monitoring — is exactly what AI risk governance lacks. This brings us to the actual market problem. You cannot short "AI kills everyone." There is no instrument for it, no index, no insurance product. Arbitrage doesn’t respect the story; it respects verifiable discrepancies, and none exist here. The gap between interior belief and market price cannot be arbitraged because there is no settlement mechanism. Bull markets do not price tail risk; they pay for it later. The date of payment is simply not on anyone’s calendar. The market’s response has been a lesson in itself. Nothing has been de-risked. AI-related tokens kept trading; technology equities kept climbing. If a double-digit probability of terminal outcomes cannot move the discount rate, then that rate is not being set by evidence. It is being set by narrative. Regulators now face a dilemma. The EU AI Act is new, untested, and written by people who have never inspected model weights. A resignation, however dramatic, gives them a public hook for rules that looked abstract six months ago. If this event moves the legislative needle, it will matter more in Brussels than in San Francisco. But do not assume regulators will solve the measurement problem. They can force disclosure; they cannot create the missing discipline that turns a subjective decimal into an auditable risk. The question that matters now is not whether AI kills everyone. It is whether the people who believe it might are willing to do the unglamorous work of building tools that make their claim testable. Resignations generate headlines. A public, reproducible risk framework would generate accountability. The next time an insider gives you a probability, ask for the model. Show me the simulations, the base rates, the causal path from architecture to extinction. If the case is real, it will survive scrutiny. If it cannot survive scrutiny, then what we witnessed this week is not whistleblowing. It is just another unsecured claim in a market that has not yet learned to demand collateral.