The Leak That Wasn't Supposed to Happen
Anthropic didn't announce its next flagship model. The internet found it anyway.
A misconfigured content management system left nearly 3,000 internal documents publicly accessible — including a draft blog post describing a next-generation model the company internally calls "Capybara" and is preparing to brand as Claude Mythos. Security researchers discovered the exposed cache. Fortune reviewed the documents before Anthropic locked them down. The company has since confirmed the model's existence, calling it a "step change" in capability and the most powerful system it has ever built.
The leak itself is embarrassing for a company whose entire brand is built on responsible, deliberate AI development. A default CMS setting that automatically made uploads public is exactly the kind of mundane operational failure that causes serious damage at exactly the wrong moment. Anthropic is in the middle of an aggressive product push, reportedly planning an IPO, and competing for enterprise trust against OpenAI and Google. Leaking your crown jewel because of a checkbox in a file upload interface is not the narrative you want.
What Mythos Is Claiming
According to the exposed documents, Claude Mythos significantly outperforms Opus 4.6 — already Anthropic's strongest publicly available model — across coding benchmarks, academic reasoning, and cybersecurity evaluations.
That last category deserves special attention. Anthropic's own framing is striking: the company describes Mythos as "far ahead of any other AI model in cyber capabilities" and warned that it "presages an upcoming wave of models that can exploit vulnerabilities in ways that far outpace the efforts of defenders."
That is Anthropic warning the world about its own model. It is an unusual move — simultaneously a capability claim and a liability hedge — and it signals that the company is aware the cybersecurity applications cut in both directions. A model capable enough to be a serious offensive tool is also a model that regulators, security researchers, and competitors will scrutinize intensely.
Anthropic says Mythos is currently in limited early access with select customers. The company is being "deliberate" about the broader rollout, which in the context of the cyber warning likely means additional safety evaluation before general availability.
A March That Was Already Extraordinary
The Mythos leak didn't happen in isolation. March 2026 has been one of the most densely packed product months any major AI lab has produced.
Anthropic shipped over 14 significant updates in roughly six weeks: computer use for Pro and Max users, Claude Code expansion to web and mobile, a 1M-token context window at standard pricing, interactive visualizations inline in the chat interface, memory for free users, and the Claude Partner Network with a $100 million commitment. Claude Code usage grew 300% since the Claude 4 models launched, with run-rate revenue up 5.5x according to the company.
Separately, Anthropic's Model Context Protocol hit 97 million monthly SDK downloads in March — up from roughly 2 million at launch in November 2025. That kind of adoption trajectory means MCP is no longer Anthropic's protocol in any meaningful sense. It belongs to the ecosystem now, which is precisely why Anthropic donated it to the Linux Foundation earlier this year.
The pace is remarkable. It is also visibly straining the company's infrastructure. Claude experienced at least five outages in March alone, with Opus 4.6 showing elevated errors even as this story developed. Shipping fast and shipping reliably are in direct tension, and right now Anthropic is winning on the former while struggling with the latter.
The IPO Timing Question
Neither Anthropic nor OpenAI is operating in a vacuum. OpenAI is reportedly preparing its own major model release — internally codenamed "Spud" — with CEO Sam Altman reportedly telling employees it will dramatically accelerate economic output. Both companies are understood to be planning IPOs later this year.
That context matters for how to read the Mythos situation. The timing of major model releases ahead of public offerings is not coincidental. Capability milestones drive valuation narratives. The leak may have forced Anthropic's hand slightly — confirming Mythos publicly rather than staying silent — but the underlying dynamic was already in motion.
This puts pressure on Google DeepMind and Meta's AI teams, both of whom are developing their own frontier models on similar timescales. A "step change" claim from Anthropic — even an accidental one — resets expectations for what the next benchmark cycle needs to demonstrate.
What This Means
The Claude Mythos reveal, however accidental, crystallises several tensions that are shaping the AI industry right now.
- For developers: The capability curve is steepening again. If Mythos delivers on its benchmarks, the gap between current production models and frontier models will widen significantly — which affects how you should be thinking about building on today's APIs versus waiting for the next generation.
- For enterprises: Reliability matters as much as capability. Five outages in a month is a product problem, not just an infrastructure footnote. Any serious deployment needs fallback logic, and Anthropic needs to address this before Mythos launch or it will undermine the narrative.
- For the security community: Anthropic's own warning about Mythos's offensive cyber capabilities is worth taking seriously. A model that "far outpaces defenders" is not a theoretical concern. Expect this to become a flashpoint in AI safety debates over the coming months.
- For regulators: The data leak itself — 3,000 internal documents exposed by a misconfigured default setting — is a reminder that even frontier AI labs have basic operational security failures. The policy conversation about AI lab transparency and disclosure obligations is about to get more concrete.
The broader picture is a company running at three speeds simultaneously: shipping consumer features at a startup pace, managing enterprise reliability at a scale it hasn't fully solved, and developing frontier models it isn't yet ready to talk about. That is an extremely difficult balance to maintain. The next few months will reveal whether Anthropic can hold all three together.
