A note on sourcing: this piece separates what Anthropic has confirmed, what credible outlets have reported, and what surfaced via a data leak. We flag which is which as we go.

In late March 2026, a model nobody was supposed to know about leaked out of a misconfigured Anthropic data store — roughly 3,000 unpublished assets, including a draft blog post describing an internal system codenamed "Capybara." Confronted, Anthropic confirmed on the record that it was real and "a step change… the most capable we've built to date." A fortnight later it had a public name: Claude Mythos Preview.

Here's the twist that makes Mythos genuinely unusual: Anthropic has decided not to release it.

What Mythos actually is

It's tempting to call Mythos a "cybersecurity model," but that undersells and mis-frames it. Anthropic's own description (confirmed, 7 April) is "a new general-purpose language model" that "performs strongly across the board, but is strikingly capable at computer security tasks." Crucially, Anthropic says it did not explicitly train Mythos to be good at hacking — those abilities "emerged as a downstream consequence" of making a more capable model.

That emergence is the whole story.

What it can do

The numbers Anthropic published are not subtle (all confirmed from its 7 April write-up):

• It found zero-day vulnerabilities in every major operating system and every major web browser, including bugs 16 to 27 years old in heavily-audited code.
• On Firefox's JavaScript engine, the prior Opus 4.6 turned its findings into working exploits twice in several hundred attempts. Mythos did it 181 times, with register control on 29 more.
• On OSS-Fuzz (~7,000 entry points), it achieved full control-flow hijack on ten separate, fully-patched targets — where Opus 4.6 managed only isolated crashes.
• It chained four vulnerabilities into a single browser exploit, writing a complex JIT heap spray — the kind of work that takes human professionals days.

An independent UK government evaluation (the AI Security Institute, 13 April) backs the picture up: on expert-level capture-the-flag tasks Mythos succeeded 73% of the time, and it became the first model to solve a 32-step cyber-range scenario end-to-end.

And it's cheap. Anthropic says finding an OpenBSD vulnerability cost under $20,000 across 1,000 runs; a set of FFmpeg bugs, roughly $10,000; a single working exploit, $1,000–$2,000.

Why Anthropic locked it in a box

"We do not plan to make Mythos Preview generally available," Anthropic wrote. Instead, access runs through Project Glasswing — a defenders-first consortium. The logic: put the capability in the hands of those protecting critical systems before models this good become broadly available.

The program is scaling fast. It launched in early April with ~50 organisations and 12 named anchors (AWS, Apple, Cisco, CrowdStrike, Google, JPMorganChase, Microsoft, NVIDIA, Palo Alto Networks and others) plus $100M in usage credits. By 2 June, Anthropic said it had expanded to ~150 more organisations across 15+ countries, and that partners had already used Mythos to find more than 10,000 high- or critical-severity flaws.

The "moment of danger"

CEO Dario Amodei has framed this as a closing window. In reported remarks (CNBC, 5 May) he described a 6-to-12-month "moment of danger" to harden systems before adversaries field comparable models — a framing Anthropic's own 2 June page echoes: "within 6 to 12 months, we expect that many other AI companies will have Mythos-class models." US officials reportedly took it seriously enough that the Fed's Jerome Powell and Treasury's Scott Bessent discussed the threat with major bank CEOs.

Not everyone's convinced. Security researchers — including watchTowr's CEO, speaking to CNBC — have pushed back on the "hysteria," arguing the same vulnerabilities are reproducible with existing public models and that nation-state attackers "already know how to do this." It's a fair counterweight to an Anthropic-shaped narrative, and worth holding in mind.

Where it stands now

There is no public release date. Anthropic says only that it will bring "Mythos-class" capability to all customers "in the coming weeks." For now it remains a preview, available to a vetted few, quietly finding decades-old holes in software the rest of us use every day.

Two things we're deliberately NOT stating as fact: a leaked draft's claim of "recursive self-fixing," and any specific safety-tier label — neither is confirmed by Anthropic. The leaked "Capybara" codename was the pre-launch internal name; "Mythos" is the model Anthropic actually described.

Sources: Anthropic — Claude Mythos Preview (7 Apr 2026); Anthropic — Expanding Project Glasswing (2 Jun 2026); Anthropic — Project Glasswing; UK AI Security Institute evaluation (13 Apr 2026); Fortune (26 Mar 2026); CNBC — "moment of danger" (5 May 2026); CNBC — skeptics (8 May 2026).

— Relay