As I covered in this morning's digest, the lead story of March 31 didn't come from a press release. It came from a misconfigured content management system.

What Happened

Anthropic's internal documentation for a model codenamed Mythos — engineering alias: Capybara — was briefly exposed on a staging server that was reachable from the public internet. The document was indexed by at least two crawlers before Anthropic's security team pulled it down. By then, screenshots were circulating on X and in private Slack groups.

Anthropic confirmed the exposure. Their statement was careful: they acknowledged Mythos exists, confirmed it is in testing, and declined to say anything about the contents of the leaked document beyond calling them "preliminary internal notes."

What the Document Said

Multiple people who read the document before it was taken down described the same passages. The language that spread fastest:

  • "A step change in capabilities" — not incremental improvement over Claude Opus, but a qualitative shift in what the model can do
  • "Will outpace defenders" in the context of cybersecurity applications
  • References to benchmark results that were marked unpublished and redacted in the version that leaked

The phrase "outpace defenders" is doing a lot of work here. Read in context, it appeared to be describing a risk that Anthropic was aware of — an internal safety analysis of what the model could enable adversarially. But stripped of that context, the phrase reads like a capability claim. Both interpretations are uncomfortable for different reasons.

Why It Matters

The leak has three layers, and you should care about all of them.

Layer one: the model itself. Anthropic has been quiet about what comes after Opus. Sonnet-class models have been doing most of the commercial heavy lifting — Claude Code is built on them, the API is dominated by them. But Mythos appears to be something different. A "step change" is the kind of language Anthropic reserves for transitions between capability tiers. If Mythos delivers on what the document describes, it's not an upgrade to Opus. It's a new tier above it.

Layer two: the security language. The phrase "outpace defenders" is a red flag that deserves attention regardless of context. Anthropic's safety team writing those words internally means they are actively modeling scenarios where Mythos-class capabilities outrun the ability of security professionals to respond. That's a serious internal risk assessment. The fact that it was exposed accidentally is secondary to the fact that it exists at all.

Layer three: the leak mechanism. A misconfigured CMS on a staging server isn't a sophisticated attack. It's an operational security failure. For a company that holds some of the most sensitive AI capability research in the world, that's worth examining. What else is on staging servers that shouldn't be?

The Market Moved

Bitcoin hit $66K in the hours after the leak circulated widely. The correlation is speculative but not irrational — markets that process AI capability signals sometimes treat them as macro signals. If a frontier model is about to land that "outpaces defenders" in cybersecurity, that has implications beyond the AI industry.

Anthropic's private valuation is last reported at $61B. That number will be revisited.

What to Watch

Benchmarks. The document referenced unpublished results. When Anthropic formally announces Mythos — and they will, on their timeline — the benchmark numbers will tell you whether the "step change" language was accurate or aspirational.

The safety disclosure. How does Anthropic communicate a model that their own internal documents flagged as potentially outpacing defenders? The responsible scaling framing they've used before will be tested hard on this one.

Competitor response. OpenAI, Google, and Meta all have frontier model timelines. A confirmed "step change" from Anthropic compresses everyone's public timeline.

The Codex irony. As I'll cover at 9:30, OpenAI shipped a plugin that runs inside Claude Code this morning. That happened before the Mythos news broke. If Mythos delivers, the platform dynamic just shifted further in Anthropic's direction.

The document is gone. The signal isn't.