▶ YouTube Summaries

• https://www.youtube.com/watch?v=UYbZo6UuMMY
• Original title: The Mythos Situation is Crazy

Forrest dissects Anthropic's rollout of Claude Mythos, the security-focused model initially declared "too dangerous" for public release and gated behind Project Glasswing, an enterprise-only initiative with Fortune 500 companies. He argues the danger framing was primarily a marketing and valuation play: the model is genuinely capable at finding vulnerabilities (OpenBSD, FFmpeg, Linux kernel, Firefox, a fraudulent wire transfer at a partner bank), but the results from real case studies like curl are more evolutionary than revolutionary, and the public release conveniently landed only after Anthropic's $965B funding round, IPO filing, and a lease of xAI's Colossus 1 compute.

The launch: Project Glasswing

Two months prior, Anthropic announced Claude Mythos as too dangerous for public release and instead rolled it out through Project Glasswing, an initiative with large enterprise customers aimed at "securing critical software." The pitch: AI models have surpassed all but the most skilled humans at finding and exploiting software vulnerabilities. Headline examples backing the claim:

• A 27-year-old vulnerability in OpenBSD, one of the most security-hardened operating systems.
• A 16-year-old vulnerability in FFmpeg, which underpins most video on the internet.
• Autonomously finding and chaining several vulnerabilities in the Linux kernel.

Forrest reads the setup as: pay millions for Mythos to find your vulnerabilities before the public gets the model that will otherwise be used to exploit them.

Why the story smelled off

Forrest's skepticism, voiced in real time back in April:

• A week before the Mythos release, the entire Claude Code source code leaked because a .map file wasn't npm-ignored; a week before that, Mythos itself surfaced in an accidental data leak. How does a model this good at security let that happen?
• Claude Code still had long-standing quality problems (the flicker bug, worse uptime than the pre-AI era, an unreliable status page) despite Anthropic internally using Mythos since February 24.
• Anthropic itself called Mythos preview a general-purpose model and ships code-review features, so "it's only a red-team model" doesn't excuse the sloppiness.
• He rejects the claim that the Claude Code team lacked Mythos access: the leaked source contained extensive Mythos-specific handling (codename Capybara) — guardrails, prompt tweaks, special behavior — that makes far more sense as first-hand integration work.

Case study: Firefox

Mozilla published a post on hardening Firefox with Mythos preview and shipped a large batch of security fixes. Anthropic's framing ("Firefox fixed more security bugs in April than in the past 15 months combined") annoys Forrest: big companies always have a backlog of known security bugs, and remediation was still done entirely by engineers. Mythos found roughly 271 bugs (about half of the fixed total; the rest were pre-existing backlog or found by humans), so the celebrated chart mostly reflects Mozilla finally allocating engineers, not the model. The honest brag would be the ~180 high severity, ~80 moderate, and 11 low severity bugs Mythos actually found — which he concedes is genuinely impressive for an effectively autonomous run (with a Mozilla-built harness).

Case study: curl

The curl creator tested Mythos and called the danger narrative "marketing": five findings, of which three were false positives, one was just a bug, and one was a confirmed low-severity CVE. His conclusion, which Forrest endorses: any project that hasn't yet scanned its source with AI-powered tooling will find a huge number of flaws with this generation of tools — Mythos and many others alike — and Anthropic may be selling an evolutionary step as a revolutionary one. Notably, Mythos also didn't find all of curl's vulnerabilities; more were found afterwards.

The theory: hype as a valuation and compute play

Forrest's timeline argument:

1. Mythos announced with fear-heavy marketing and exclusive enterprise access, driving big clients to spend more, publicly.
2. Anthropic was simultaneously raising its final round at a $965B valuation (up from $380B ~3.5 months earlier) and then filed for an IPO.
3. Public rollout of the "insanely dangerous" model was announced only after the raise, the IPO filing, and an agreement to lease all 300 MW of xAI's Colossus 1 compute.

His read: Mythos was likely ready earlier, but Anthropic lacked the compute to serve everyone, so scarcity was rebranded as danger — pumping the valuation along the way. "But that's just a theory."

Takeaways

• Mythos is a real, capable security model — comparable claims could be made for GPT 5.5 or Opus 4.8 — but the "destroys all software" framing was overblown fear marketing.
• Treat vendor case-study charts skeptically: who found the bugs, who fixed them, and what backlog already existed are separate questions.
• AI-assisted vulnerability scanning is now table stakes; unscanned codebases will yield many findings regardless of which frontier model you use.