Claude Mythos: Anthropic's Top AI Leaks, Cyber Threat

Claude Mythos — Anthropic's next-generation AI model described internally as the company's most capable build to date — leaked online this week after draft materials tied to the company's blog were left in a publicly accessible data store, triggering a swift market reaction that erased billions from cybersecurity stocks on Friday.

What Is Claude Mythos and How Did It Leak?

The model's existence was confirmed by Anthropic after Fortune first reported that unpublished files — including what appeared to be a staging blog post — had been discoverable in an unsecured content cache. Anthropic restricted access to the data store after being notified, attributing the slip to a misconfiguration of its CMS tooling. Human error, they called it. For a company handling arguably the most sensitive AI development on the planet, that framing is going to age awkwardly.

In an archived development page reviewed independently, Claude Mythos was described by Anthropic as "the most powerful AI model we've ever developed" — and the company wasn't being modest about what that means. According to those materials, Mythos sits in a tier above the Opus model line entirely, described as "larger and more intelligent than our Opus models — which were, until now, our most powerful." The name itself, Anthropic wrote, was chosen "to evoke the deep connective tissues that link together knowledge and ideas." Poetic. Also alarming, depending on which sentence you read next.

Anthropic confirmed the broad strokes in a statement: "We're developing a general purpose model with meaningful advances in reasoning, coding, and cybersecurity. Given the strength of its capabilities, we're being deliberate about how we release it." The company said it has been working with a small group of early access customers to test the model ahead of any wider rollout.

We consider this model a step change and the most capable we've built to date.

The Cybersecurity Warning Buried in the Leak

Here's the part that actually moved markets. According to the leaked draft materials, Mythos scored "dramatically higher" than Claude Opus 4.6 on benchmarks for software coding, academic reasoning, and — critically — cybersecurity offense. That last category is where things get uncomfortable.

The cybersecurity stocks sell-off on Friday wasn't just investors spooked by a new AI product announcement. It was a specific reaction to language in those leaked documents that described Mythos as "currently far ahead of any other AI model in cyber capabilities" — and then went further, warning that the model "presages an upcoming wave of models that can exploit vulnerabilities in ways that far outpace the efforts of defenders." Anthropic wrote that themselves. That's not a researcher's speculation — it's the company's own internal framing of what they've built.

Palo Alto Networks (PANW) fell roughly 7% on Friday. CrowdStrike (CRWD) dropped around 6.4%. Zscaler (ZS) declined approximately 5.8%, and Fortinet (FTNT) slipped about 4%, according to data from Yahoo Finance. The message from markets was unambiguous: if the most advanced AI model on the planet was built with cybersecurity attack capabilities baked into its benchmarks, and its creator is releasing it to a curated list of early testers — the existing defense software stack starts looking a lot more fragile.

A Second-Tier Model, and What Comes After Mythos

The leaked documents didn't stop at Mythos. Anthropic's materials also referenced a second, follow-on model internally codenamed "Capybara" — described as a version two in the same tier, positioned above the current Opus line as well. That detail has received less attention than the initial Mythos headline, but it matters: Anthropic appears to be building out an entirely new capability tier, not just refreshing an existing model family.

Because of the cybersecurity risks it assessed in its own documents, Anthropic said it plans to release Mythos cautiously, starting with organizations working on cybersecurity defense specifically. That makes intuitive sense — if you've built a model that can find and exploit vulnerabilities faster than defenders can patch them, handing it first to the people running defense operations is the responsible call. Whether those early-access guardrails hold once the model starts circulating more broadly is a different question entirely.

The financial blowback echoes what happened in February, when Anthropic unveiled Claude Cowork — an AI system for automating complex workplace tasks like contract review and compliance. That announcement triggered a broad sell-off across software and professional-services companies that erased roughly $285 billion in market value, as investors began repricing the structural risk of foundation model providers moving up the stack into enterprise software territory.

That's a polite way of saying if Anthropic can build a legal workflow tool in-house, what's stopping them from doing the same for finance, procurement, or HR?

What Does the Mythos Leak Mean for the AI Arms Race?

The leak itself — accidental as it was — landed in a particularly charged week. Nvidia CEO Jensen Huang had just told Lex Fridman's podcast that "I think we've achieved AGI." Two days later, the ARC Prize Foundation released ARC-AGI-3, the most rigorous AGI benchmark in current use, and every frontier model scored below 1%. Google's Gemini led the pack at 0.37%. OpenAI's GPT-5.4 came in at 0.26%. Anthropic's own Claude Opus 4.6 managed 0.25%.

So the AGI claims and the benchmark reality are living in separate zip codes right now. But that context makes the Mythos leak more interesting, not less. Anthropic isn't claiming AGI — it's claiming a model that scores "dramatically higher" on coding, reasoning, and cybersecurity than anything currently deployed. Whether that closes the gap on benchmarks like ARC-AGI-3 or opens entirely new threat surfaces is exactly what those early-access testers are presumably trying to figure out.

A federal judge also blocked the Pentagon — on Thursday — from labeling Anthropic a supply chain risk, ruling the government's campaign against the company violated First Amendment and due process rights. U.S. District Judge Rita Lin issued the preliminary injunction from the Northern District of California. The Pentagon fight isn't over, but Anthropic walked out of that ruling with some breathing room. And now, with Mythos out in the open, the real negotiation begins — not between Anthropic and the government, but between Anthropic and every CISO who just watched their vendor's stock drop 7% in a single afternoon.