On a Tuesday in early September 2026, Jacob Coxon did something that employees at frontier AI labs rarely dare: he resigned in public, and he resigned with a warning that belonged to fiction. Three years as a pre-training researcher at Anthropic, the company that built its entire brand on the words “safety first,” gave him the credibility to say what he said. Superintelligence is coming, he wrote—not a hypothesis, not a hedge—and by the end of this decade, it will kill everyone.
The words landed like a brick dropped into a still pond. Evan Hubinger added a number to the dread: more than ten percent probability of extinction. Open letters followed. More than 1,100 employees, including chief scientists, signed calls to slow frontier development. Former OpenAI figures stepped forward to warn about rogue AI. The United Nations human rights chief, Volker Türk, felt compelled to speak publicly about the stakes. Meanwhile, neither Anthropic nor OpenAI paused a single training run. Not one. They issued statements. They thanked departing employees for their contributions. The GPUs kept humming.
From my seat in Cape Town, watching the crypto market enter this new bull phase with all its familiar euphoria, I have seen this exact movie before. Different acronym. Different technology. Same uncomfortable shape: a small group of people with deep insight grows convinced that the ship is heading for the rocks, while the people at the helm keep adjusting the deck chairs and the music never stops. Tracing the code back to the conscience behind it has always been my discipline. This story gives us a chance to practice it on a much larger stage.
Anthropic was supposed to be different. It was literally built as a counter-narrative to OpenAI's scorched-earth march toward the frontier. Founded by safety researchers who broke away from OpenAI precisely because they felt commercial velocity was drowning alignment work, Anthropic promised a different trajectory: safety research first, models second. The corporate structure itself was meant to function as a guardrail for conscience. For a while, the story held. Constitutional AI frameworks. Red-team protocols. Written commitments to responsible scaling. Enterprises loved the narrative. Regulators appreciated having a responsible player to point to while they drafted the first generations of AI law. Anthropic became the adult in the room, the lab you could bring home to your parents.
Then the pattern starts to smell familiar. In February 2026, safety lead Mrinank Sharma walked out. Then Coxon surfaced. Then the open letters revealed a startling internal consensus: the very people building the models believed the models might end us. What does a cluster of high-level departures mean right before a market cycle cracks open? Ask anyone who watched the centralized exchanges of 2021. It means the insiders with access to unvarnished information no longer believe the visible corporate narrative.
Coxon's description of the Anthropic dilemma is perhaps the most precise sentence in this entire messy episode: the company understood the risks but remained trapped in the race to be first. There is always one more model before safety. One more push to arrive. That is not the same thing as an organization lying. It is something more common, and scarier: an organization incapable of stopping itself.
For those of us who lived through the collapse narrative of centralized custodians, this sequence is already written. No amount of narrative or marketing control can substitute for verifiable infrastructure when the core team keeps moving the product forward at a speed that governance cannot sustain. And the first casualty of speed is usually the truth.
Let me say something unpopular about safety teams, whether they guard ERC-20 protocols or colossal neural networks: the people inside those teams who are closest to the code, and therefore best positioned to see the catastrophe coming, are often treated as inconvenient. In 2017, during the ICO boom, I spent four months auditing ERC-20 token standards for three promising Cape Town projects. I identified critical reentrancy vulnerabilities in two of them, vulnerabilities that would have allowed attackers to drain investor funds in a single malicious transaction. My public documentation on GitHub saved approximately $45,000 in potential losses. A drop in the ocean compared with what an AI alignment failure could consume, but the pattern was identical.
When I flagged those vulnerabilities, I met resistance rather than gratitude. As one of the few women in the local crypto circle, I was used to being doubted. The projects took my feedback, thanked me politely, claimed they would iterate quickly, and then shipped ahead of the patch cycle anyway because being first to market mattered more than being safe to market. Both projects later collapsed. Their investors never recovered their funds. The security warnings had been accurate, precise, and publicly documented. None of that mattered against the gravity of the incentive structure.
Now multiply that story by the scale of the entire human species, and you will begin to glimpse the reality of 2026. High leverage. Dangerously small feedback loops. Confident narratives louder than real verification. This is why Coxon's resignation letter is not merely a story about one failing lab. It is a universal pattern, and it deserves the technical scrutiny we usually reserve for protocol vulnerabilities.
Start with the structural flaw: safety teams without safety authority. Anthropic's alignment personnel, the very group tasked with slowing dangerous deployment, did not appear to have the authority to stop it. Coxon worked in pre-training; Sharma worked at the safety frontier. Neither held the decision rights to pause the race. That simple observation completely reframes the open letters. These were not employees expressing concerns through established channels and being ignored. These were employees who understood that the established channels were designed to absorb concern without ever converting it into action.
This recalls a narrative we see in our own industry, the so-called liquidity fragmentation problem. The argument says liquidity pools split across chains and protocols create inefficiency, so we should build megastructures to consolidate them. I have always maintained that this is a manufactured crisis, a story told by those who sell consolidation services. In Anthropic's case, the safety narrative is similarly manufactured at the product level. Constitutional AI does not mean the model is safe. It means the model is programmed to behave as if it has internalized a set of rules. The pre-training personnel now tell us these rules cannot be guaranteed. The constitution is a persuasive interface, not a verified property.
In a distributed network, if a validator or a small group of nodes defects, the rest of the network can fork away and continue functioning. That is the entire argument for decentralization. But inside Anthropic's control architecture, the people with the highest epistemic insight into alignment have no fork. They can either stand up on the way out, as Coxon did, or stay inside and pretend not to see what they see in the training runs. Departure becomes the only governance tool available. And the market needs to understand what that means: when the only governance tool left is ejection, the system itself has already failed.
Now follow the market signal. We know from our own history that exchange-based distribution was once a golden pipeline. Binance Launchpad returns fell from 100x to 10x over time, proof that traffic monetization decays as the market matures and the marginal participant grows wiser. The same decay curve now applies to the safety premium in AI. For years, Anthropic charged a premium in enterprise contracts and policy influence based on its identity as the safe AI company. The moment the market computes that safety is a slogan rather than a technical property, that premium collapses. Each resignation is a discount applied to the brand.
I have watched this movie before, in a different costume. FTX built a brand around responsible management and regulatory engagement. Its customers believed the audited financials, believed the celebrity endorsements, believed the polished narratives about doing good while doing well. The collapse was not caused by a lack of stated values. It was caused by a lack of verifiable controls. Anthropic does not have to be a fraud to follow the same trajectory. It only needs to be convinced that its own marketing is true while its internal operations say otherwise.
The letters reveal another layer that deserves skeptical attention. When more than 1,100 employees sign documents saying slow down, while management continues the march with no meaningful pause, what exactly is being accomplished? One possibility is that the employees are sincerely alarmed and using every tool available. That is generous and probably partially true. The second, more cynical possibility is that these letters function as regulatory hedging. By leading the public discourse on restraint, the large frontier labs position themselves to shape the rules in their favor. They become the reasonable voices that regulators invite to the table when drafting new laws. They set the terms of the debate, and in doing so, they make it harder for any future competitor to disrupt their dominance.
Our own industry experienced this exact mechanism. The Markets in Crypto-Assets regulation, MiCA, is presented as the first clear legal framework for crypto in the European Union, bringing long-awaited certainty. In practice, the compliance costs and reserve requirements are so onerous that small projects will be crushed under the burden while the large centralized players absorb the expense and secure their dominance. Regulation is not always a shield for the vulnerable. Sometimes it is a moat around the castle.
The parallel writes itself: when the biggest labs call for a pause, they are the only institutions with the compute capacity, the talent pools, and the capital reserves to weather that pause comfortably. A startup that emerges next year with a novel alignment approach would face existential risk if a global cap arrived, while OpenAI and Anthropic would simply extend their head start. There is no version of this story where slowing the frontier meaningfully reduces market concentration. In fact, the pause looks increasingly like a regulation designed to protect giants, signed by the most conscience-stricken insiders as a form of moral insurance.
I do not want to be entirely cynical about those who signed. As someone who facilitated a developer mental health support group called Code & Conversation during the brutal 2022 bear market, I understand that humans can be simultaneously sincere and trapped. I sat through fifty one-on-one sessions with developers whose portfolios had lost eighty percent of their value, many of whom felt personally responsible for failures that were systemic in origin. The emotional labor of that period taught me that individual conscience is real, but it operates inside incentive structures that overwhelm it. If I worked inside Anthropic and genuinely believed my organization was about to end the human story, I would stop using internal escalation routes that clearly do not function. I would consider my public departure a strategic act, not merely a moral one.
This brings us to the structural need for verifiable alignment. Blockchain, at least in its open forms, offers continuous auditability. The chain is visible. The records are shared. Independent validators can check the state. Centralized AI labs offer nothing equivalent. There is no public node that external observers can query to verify whether alignment tests are actually being run or whether an evaluation for dangerous capabilities has been suppressed. The opacity is not incidental. It is organizational design.
If there has ever been a moment in our technological history to re-examine the philosophical foundations of verification, this is it. We are about to cross a threshold where a tiny fraction of institutions fully understand the technology that will make consequential decisions for the rest of humanity. The centralization of intelligence is a challenge that makes the concentration of the monetary system look like a small governance problem in a summer thunderstorm.
What would a decentralized and audited alignment environment look like? In the best possible world, we would have public, externally verified measurements of frontier model capability. Independent red teams with hardened credentials, operating outside the control of model creators, would hold the keys to evaluation. Model weights would be structured to permit third-party scrutiny without triggering an arms race between labs. And we would fight constantly to prevent safety-verification culture from being captured by the same marketing teams that sell safety as an adjective rather than a proof.
I have touched this possibility in my own work. In 2025, I collaborated with a global team of fifteen researchers on a project integrating decentralized identity protocols with AI verification systems. We wanted users to prove the origin of digital content without revealing personal data. The technical challenge was genuinely satisfying: we needed the privacy guarantees of decentralization and the proof requirements of cryptography, simultaneously. After months of iteration, we piloted the framework with 5,000 users and prevented approximately 2,000 instances of identity fraud. That success was modest in scale but profound in its implication. It proved that open, community-driven design can still deliver credible security properties in an age dominated by closed mega-labs. Code, and only code, should verify code. The people who write the rules should not be the same people who decide whether the rules are being respected.
Open source is not a license; it is a promise. Yet the deepest lesson from Anthropic's exodus is darker. The laboratory that most loudly promised alignment is now the laboratory that cannot retain its alignment researchers. Each public departure is a discovered vulnerability in the corporate claim to responsibility. The exploit is not in a smart contract. It is in the gap between what organizations say about themselves and the accountability instruments they actually deploy.
Let us consider the market that is emerging from this exodus. Coxon's resignation might mark the beginning of a professional migration pattern, from well-compensated lab alignment teams toward independent research, watchdog work, and public auditing. In crypto, we have already watched dozens of security engineers leave centralized institutions to start independent protocol audits and public security research groups. The market adapted to that shift. It will adapt to the same shift in AI. The defection pathway is not purely heroic. It is also a market outcome. Anthropic's specialists are among the best in the world at articulating intelligence risk. If the lab ignores their concerns, those individuals suddenly gain larger personal platforms and more valuable consulting contracts. Coxon's market value likely increased the moment he resigned.
That inverts the logic of control. Instead of keeping dissenters quiet, the lab's behavior converts them into well-compensated, highly credible critics. The cathedral is hollowed out from within, and each departure produces a new source of institutional critique with the authority of lived experience. This dynamic makes the coming period extraordinarily dangerous for centralized labs and arguably stabilizing for the rest of us.
Did I ever consider resigning from this ecosystem in dark moments? Yes. In 2022, when the bear market erased portfolios and trust in our industry reached its low point, I felt the same pull toward escape. What anchored me was not any particular protocol or token. It was the community that gathered weekly in Cape Town to learn, to audit failed projects, to turn despair into structural lessons. We were building bridges, not just blocks, between people. That experience taught me that translating risk awareness into community resilience is more valuable than individual escape. Maybe that is what we must do with the AI safety exodus: treat it not purely as a crisis of one company, but as an urgent argument for building an independent, decentralized verification class that does not depend on corporate sponsorship for its existence or its funding.
But here is the contrarian angle, and I offer it because my own community needs to hear it. If you believe decentralization automatically solves AI alignment, you have not thought carefully about malicious use cases. Open-sourcing an advanced superintelligence, in a world where an aligned version remains elusive, could cross a moral red line that cannot be uncrossed. In crypto, permissionless innovation is generally welfare-improving because a bug in a protocol affects a bounded pool of value. In AI, a bug in an open model could be a global catastrophe. If a lab open-sources the weights of a frontier system without verified safety properties, every adversarial group on earth receives raw material to modify and exploit. The world would not be safer. It would be exposed in a way that our industry rarely considers.
This asymmetry is why I cannot simply wave the banner of decentralization as a silver bullet. We need something more precise: verifiable delegation. A set of independently audited standards that constrain harmful behavior at the protocol level before any open invocation is considered. This is an extremely thin line to walk. There is no global government that can enforce transparency, and no protocol that can be enforced globally without introducing the very centralized authority that many of us reject. For the first time in my career, I find myself admitting that the old answers of just open everything or just centralize authority are both inadequate to the scale of this question.
Perhaps the answer requires a form of distributed accountability that does not yet exist. The departing researchers have given us a gift, even if they do not see it that way. They have shown us what happens when the people closest to a technology lose faith in the institutions building it. That loss of faith is a signal. The architecture of control that depends on keeping warnings inside is already obsolete. What replaces it must be built in the open, before the next training run completes.
Every line of code is a hand extended in trust. The alignment community is telling us that trust in centralized institutions is no longer sufficient. We need checks that do not depend on the goodwill of the checked. We need measurements that survive the departure of the people who created them. We need a governance layer designed before the emergency, not during it.
Education is the only true decentralized currency, and this time the curriculum is generation-scale. The researchers who resigned have become teachers by example. They have demonstrated that knowing the risk and naming the risk are two different acts, and that naming it may cost you your place inside the system. We should honor that teaching by building the verification infrastructure that makes such resignation less necessary.
Artists own their pixels; we just hold the keys. In the age of synthetic intelligence and decentralized ledgers, the equivalent principle is that communities own their future; institutions merely hold the authority. When institutions lose the trust of the people who understand them best, the authority should return to the community. The question is whether we have built the mechanisms to receive it. We build bridges, not just blocks, between people, and right now the most important bridge is the one connecting independent technical verification to public accountability. Let us hold our keys with integrity while there is still time to build that bridge.
