Mythos AI achieved 595 crashes in pre-release safety evaluation
On Friday, April 10, executives from Canada's six largest banks and Desjardins Group sat in a room with representatives from the Bank of Canada, the Office of the Superintendent of Financial Institutions, the Department of Finance, and the Canada Deposit Insurance Corporation. The topic was Claude Mythos.
This was not an emergency activation of some crisis protocol. The Cyber Resilience Forum Group, the body that convened the meeting, was careful to make that distinction. The Globe and Mail reported that the CFRG can still hold situational awareness meetings at the request of its members — the chair describing it as: hey guys, we need to pay attention, there is something going on, let's get together and talk about this. That is what this was.
But the fact that it happened at all, days after the United States convened its own meeting with Treasury Secretary Scott Bessent, Fed Chair Jerome Powell, and the CEOs of Bank of America, Citigroup, Goldman Sachs, Morgan Stanley, and Wells Fargo, tells you something real about how financial regulators are thinking about this model.
A model that finds bugs old and new
Anthropic's own red team report, released alongside Mythos, is unusually detailed for a pre-release safety evaluation. The numbers are the kind that get attention in the rooms where these decisions get made. Mythos Preview achieved 595 crashes at tiers 1 and 2 — the lower end of the harm spectrum — a handful of crashes at tiers 3 and 4, and full control flow hijacks on ten separate, fully patched targets. Opus 4.6, for comparison, turned vulnerabilities it found in Mozilla's Firefox JavaScript engine into working JavaScript shell exploits two times out of several hundred attempts. Mythos did so 181 times.
The model also surfaced a now-patched 27-year-old bug in OpenBSD. Anthropic says finding it cost approximately $20,000 per discovery campaign. The specific model run that surfaced the flaw cost under $50.
Those are the numbers that landed on regulators' desks. The CFRG meeting was hastened by Mythos's release but not triggered by a specific incident. The question on the table was anticipatory: what happens when a model this capable is in wider use, and who bears responsibility for what it finds?
The Canadian forum, and who did not attend
Alexis Corbett, the chief operating officer of the Bank of Canada, chairs the CFRG. She was in the room. Neither Bank of Canada Governor Tiff Macklem nor Senior Deputy Governor Carolyn Rogers attended — a detail that matters, because it suggests this was an operational-level convening rather than a political one. The governor-level absences are a signal that no one is ready to call this a crisis. They are also a signal that the operational staff are being left to manage the early assessment without being told to calm down.
The participants included the six major Canadian banks and Desjardins Group alongside the full roster of federal financial regulators. That is a broader institutional coalition than showed up in the United States, where the initial meeting was convened through Treasury and the Fed rather than through a standing industry forum.
The Glasswing response
Anthropic did not release Mythos without a response prepared. Project Glasswing, announced alongside the model, commits up to $100 million in usage credits for Mythos Preview and $4 million in direct donations to open-source security organizations. The launch partners are a list that reads like a cross-section of critical infrastructure: Amazon, Apple, Broadcom, Cisco, CrowdStrike, Google, JPMorganChase, the Linux Foundation, Microsoft, Nvidia, and Palo Alto Networks. More than 40 organizations are participating. The structure is a defense coalition assembled in advance of the threat it is meant to address.
The Bank of England is next. Mythos will be on the agenda for the BOE's Cross Market Operational Resilience Group and CMORG AI Taskforce meetings within the next two weeks.
What the pattern tells us
The global response to Mythos has followed a consistent sequence. The United States moved first — Bessent and Powell summoning Wall Street within days of the model's appearance. Canada followed several days later through the CFRG. The United Kingdom is next. Each convening carries the same underlying question, phrased differently by each institution: what are the obligations of a lab that releases something this capable, and what are the obligations of the institutions it might affect?
The answer that Glasswing offers is: partnership, not prohibition. Anthropic is not asking regulators to block the model. It is asking them to be inside the response. The 40-plus organizations in Glasswing are not bystanders — they are in the room.
What remains unclear is whether that is sufficient. The CFRG meeting on Friday was described as situational awareness. There were no public commitments, no formal asks, no regulatory actions flagged. The banks came, listened, and went back to their own internal assessments.
That is a different kind of signal than a public statement would be. Situational awareness means: we are watching, we are talking to each other, and we are not ready to say what comes next.
The story of Mythos is not just about what the model can do. It is about the pace at which the institutions it might harm are trying to organize around it. The Canadians moved faster than they might have. The Americans moved faster than they usually do. The British are next. None of them have said what they will do if the answer to the question they are asking turns out to be worse than they think.
We will find out together.