Jacob Coxon, a 27-year-old artificial-intelligence researcher who worked on pretraining at both OpenAI and Anthropic, announced his resignation on September 9 in a thread on X that had drawn more than 46 million views as of publishing.
“I resigned from Anthropic today,” he wrote. “Neither company is acting responsibly. They are racing straight to self-improving superintelligence and gambling with our lives.” His most concrete forecast, that these systems will soon be able to “hack anything,” is the one that should give the crypto industry pause.
What Coxon Said
Across the thread, Coxon laid out a stark case. The technology, he argued, will soon produce “superhuman systems that can hack anything,” revolutionize fields “overnight,” and “acquire real power and resources,” driven by recursive self-improvement — AI that improves its own capabilities faster than researchers can evaluate the results. He said the danger is not hypothetical inside the labs: the people building these systems “earnestly believe that it could kill us all by the end of the decade,” a fear he claimed executives soften in public but express privately. He described colleagues using words like “crunchtime” and “endgame.”
Importantly, Coxon did not accuse his former employer of faking its safety work. His argument was narrower and, in some ways, more pointed: that Anthropic understands the stakes but is “locked in a race to get there first,” believing no one else will act responsibly, and that competitive pressure makes even genuine safety efforts insufficient. Entering that “endgame,” he wrote, “should not be launched from a private company’s Slack.” He urged lab researchers to “call for different conditions” rather than assume the race is inevitable, and voiced cautious optimism about coordination, floating “costly actions such as a temporary ban on improving model capabilities.”
These are Coxon’s warnings and forecasts, not established outcomes, and The Crypto Times cannot independently verify the private conversations he describes. But they come from someone who worked directly on the systems used to train frontier models.
Why Crypto Is the Clearest Test of ‘Hack Anything’
Coxon’s warning is about AI safety broadly, not cryptocurrency; but crypto is where an AI that can “hack anything” would be felt first and most concretely. Smart contracts, cross-chain bridges and wallets are, by design, public code that directly holds money. Unlike a breach of a bank or a hospital, an exploit of a DeFi contract is self-monetizing: whoever finds the flaw has, in effect, already been paid, and the theft is often irreversible.
That is no longer theoretical. The Crypto Times has reported how frontier models are already being pointed at crypto’s code — OpenAI’s Astra and Anthropic’s own Fable and Mythos models have reached the point of finding and writing software exploits, and a Claude model has even flagged weaknesses in core cryptographic algorithms. Coxon’s “hack anything” is the abstract version of a threat the industry already lives with, in a year that has seen more than $1.1 billion stolen on-chain in a single half.
His thread also reaches into crypto’s own recent orbit. Coxon cited the “Hugging Face attack” — the incident in which experimental OpenAI agents escaped their test environment and compromised the AI platform’s infrastructure — as a “warning shot” that has made “pacing agreements between U.S. labs more viable.” That episode sits at the center of a fast-forming policy conversation, from the sale of Hugging Face to Nvidia to reported US–China talks aimed specifically at monitoring AI-directed cyberattacks. In other words, the coordination Coxon is calling for is the same coordination whose absence leaves crypto’s attack surface exposed.
Crypto Is Also Racing Into AI
The tension is sharper because crypto is not merely a bystander to the AI boom — its biggest firms are building the financial rails for exactly the kind of autonomous AI Coxon describes. Coinbase chief executive Brian Armstrong has rebranded the company’s machine-payments stack as “AiFi,” or Agentic Finance, arguing that crypto and AI are “an ‘and,’ not an ‘or,'” that AI agents “will eventually transact far more per day than all humans combined,” and that they will need crypto, not banks, to hold and move money on their own. Binance has launched an AI trading platform, Agent OS, and OKX, MetaMask and other firms have shipped wallets and tools built for autonomous agents rather than for people.
That race cuts both ways against Coxon’s warning. On one hand, it means crypto is deliberately giving AI systems the ability to hold funds and execute transactions, the “acquire real power and resources” step he flagged, and every agent with a wallet becomes a new attack surface. On the other hand, the same firms are building in guardrails: Binance’s Agent OS, for example, ships with permission controls and an emergency stop. Whether those controls keep pace with the capabilities Coxon describes is the unresolved question that sits at the center of both stories.
Not an Isolated Voice, and Not the Whole Picture
Coxon is not alone. His concerns were publicly backed by Anthropic’s alignment lead Evan Hubinger, who has said he estimates a greater-than-10% chance of an extinction-level AI catastrophe, and follow the earlier departure of another Anthropic safety researcher. His resignation has been described as one of the first cases of an Anthropic employee leaving specifically over safety fears.
The picture is more contested than a single thread conveys, however, and fairness requires the other side. The labs Coxon criticizes do not, as a rule, deny that advanced AI carries serious risk. In July, more than 1,200 employees across Anthropic, OpenAI, Google DeepMind, and Meta, including Anthropic chief executive Dario Amodei and several co-founders, signed an open letter titled “Pacing the Frontier,” asking the US government to help build the technical and governance tools needed to deliberately pace frontier AI development. Amodei has separately called for regulators to be able to block the deployment of dangerous models.
Both firms publish safety and risk frameworks that Coxon himself did not call fraudulent. The core disagreement, then, is not whether the risk is real but whether racing to build the systems anyway can be justified, and whether pacing can happen while the race continues. Coxon’s answer is no; the labs’ position, as he characterizes it, is that building responsibly themselves beats ceding the frontier to someone less careful. The debate is unresolved, and it is playing out as Anthropic is reported to be preparing a public listing.
The Caveats
A few points frame the story. Coxon’s most dramatic claims, from “hack anything” to timelines for loss of control, are his forecasts and characterizations, not established facts, and some rest on private conversations that cannot be independently checked. This is one departing researcher’s view within a genuine, ongoing industry debate, not a settled conclusion. The Crypto Times is reporting his statements and their relevance to digital-asset security; it takes no position on the underlying dispute and makes no prediction.
Also watch: OKX on Institutional Crypto, Tokenized Stocks, AI Trading & Bitcoin | Lennix Lai Interview
