No Proof of Concept: Jacob Coxon, Anthropic, and the Unaudited Endgame

Jacob Coxon did not leak a model. He did not open a red-team report. He did not publish a proof-of-concept on his way out of Anthropic. He resigned and posted a thought: a self-improving superintelligence is coming before the decade closes, and it could kill everyone. A week earlier, another researcher put terminal probability above ten percent. A safety lead had already left in February. The UN human rights chief is now using the phrase existential risk. None of these people produced a reproducible artifact. Yet the market absorbs all of it as price discovery. The event reads as a single data point in a chain: Jacob Coxon leaves Anthropic, cites survival-level risk, joins a growing queue of resignations and open warnings. More than 1,100 people have signed industry letters asking for a slowdown. Senior executives and chief scientists are on the list. The two leading labs continue racing. The public receives all of this as evidence because the source is internal. The article that assembles the timeline sits on a crypto media platform, not an AI audit desk, and its evidence is a collection of posts. That does not make the warnings false. It makes them unaudited. Coxon's core claim reduces to four statements. Current pretraining trends hide a capability jump. That jump enables a self-improving superintelligence. That mode will exploit software, finance, and physical resources. The result ends human governance. The first claim is possible from inside pretraining. The second is a research goal, not an observed event. The third is a threat model. The fourth has never been modeled by an independent third party. In my line of work, claims of this shape are unaudited. I spent six weeks auditing 0x v4 smart contracts in 2020. The worst bug was hiding in a conditional ordering of ERC-20 allowances that could be gamed by a transaction sequence. I reported it only after writing a reproduction and a patch. Coxon's warning has no reproduction. No model evaluations. No benchmarking trace. No internal safety dataset. A claim that depends entirely on position is a claim without necessary context. His three years in pretraining means he may have seen an internal capability jump. But access to a signal is not proof of a mechanism. That absence matters because Anthropic has branded itself as the safest fork of frontier AI. Continuous exits from security-adjacent roles spend that brand equity. Mrinank Sharma was a safety lead. Jacob Coxon worked in pretraining for three years. When internal safety voices leave, recruiting costs rise, enterprise trust falls, and the remaining team's reviews become harder to defend. All of this is commercially measurable and expensive. Anthropic's differentiation from OpenAI is narrowing. OpenAI can point to an adventurous brand; Anthropic cannot afford to lose its safety claim at the same time that it loses the people who verified it. The harder economic fact is that stated values compete with a race. Public letters with 1,100 signatories ask for slowdown, while the labs keep training at full speed. That is not necessarily hypocrisy. In a startup valuation, expected AGI arrival is the asset. A deliberate pause allows a competitor to break the threshold first and destroys the capital stack. Financial floors bind where ethics do not. Public letters function as hedges: they give regulators evidence of good faith without touching the accelerator. This is selective disclosure with a rational camouflage. Now consider the second-order effect. UN officials using terms like existential risk turn a lab dispute into a policy signal. When global institutions begin to speak in terminal language, states have a justification to act. The exact mechanics of the warning matter less than the timing. A resignation at the right moment can become the human face of a governance argument. That is why this is not purely a technology story. It is an institutional strategy event. The convergence of two industry letters in 2026 points to a coordination layer across companies. People inside different labs seem to be aligning policy language before regulators force alignment by law. The uncomfortable insight is that some frontier executives may silently need this pressure. The two leading labs face a prisoner's dilemma. If they genuinely believe that first arrival is dangerous, neither can stop unilaterally. The only outside actor with the power to enforce a pause is the state. The most efficient move is to make internal risk loud without making corporate policy explicit. A resignation on social media turns a company embarrassment into a regulator's justification. Coxon becomes a signal mechanism. The lab loses talent and gains political cover. No evidence proves this coordination, but the incentive structure explains a pattern that otherwise looks irrational. There is also a false precision in the probability. Saying the terminal risk is above ten percent sounds exact. Precision is not accuracy. In cryptography, setting the wrong false positive rate causes an availability failure. In oracle risk, assigning a probability without a distribution creates false trust. The article does not include Hubinger's assumptions. It gives no posterior distribution, no confidence interval, no stress test. Without inputs, an output cannot be audited. Anyone who has stress-tested a lending protocol knows that a number with no scenario is not a result; it is a hope. Code does not lie, but it often omits context. Here there is no code. The omission carries the real meaning. The hidden context is not that AI will end the world. The hidden context is that internal risk committees have become ceremonial, and some researchers no longer believe the company will adjust. That conclusion does not require a physical apocalypse. It only requires a governance breakdown. Read Coxon's message that way, and it changes from a scientific prediction to an auditor's refusal to sign. The standard is a ceiling, not a foundation. Coxon's testimony is a ceiling. It marks what one person could see from inside a company. It is not a foundation for a pause because a pause is an economic, political, and strategic decision. What would be a foundation? A formal threat model published by the company. Parameters, dependencies, attack paths. Independent auditors running the same scenario. That would be priceable. A resignation letter is a sentiment index. Parsing the chaos to find the deterministic core, I find one deterministic fact. No third party has verified the model risk. The first frontier lab to publish a verifiable existential threat model will set the new standard. Until then, every exit and every unsupported probability is a form of noise in an information-poor market. The policy question is not whether Jacob Coxon is brave or right. It is whether governance will rest on testimony or on repeatable verification. Decentralized finance made that choice poorly. Projects claimed security after audits, and the market learned that an audit is not a proof. The AI frontier is approaching the same lesson. Let the warning be published as a specification. Let the reproduction be attached. Until then, an unverifiable endgame deserves attention but not certainty.