A data leak just exposed Anthropic's most powerful AI model — Claude Mythos — before Dario Amodei was ready to reveal it. Here is what the leaked documents reveal about the model that even the Pentagon couldn't get its hands on.
What happens when the most powerful AI model ever built leaks into the open before anyone is ready?
That is not a hypothetical. On Thursday, March 26th, cybersecurity researchers at LayerX Security and the University of Cambridge independently found what Anthropic had accidentally left sitting in a publicly accessible, unsecured data cache — a draft blog post announcing the existence of Claude Mythos, a model the company itself describes as "by far the most powerful AI model we've ever developed." Anthropic had not announced it. The model was not available for public use. And yet there it was: three thousand unpublished assets, including what appeared to be the full product launch announcement for a model that poses, in Anthropic's own words, "unprecedented cybersecurity risks."
Dario Amodei had a plan. The leak was not part of it.
The story of Claude Mythos is not just about a model. It is about the collision of raw technical ambition, institutional paranoia, a Pentagon contract war, and a rivalry between Amodei and OpenAI CEO Sam Altman that has become one of the defining storylines in the technology industry. The Rundown AI and Superhuman AI reported on the Pentagon clash last week, but both newsletters missed what the leaked documents actually reveal about Anthropic's next move. TLDR AI noted the data breach without digging into the full context of what Mythos changes about the competitive landscape. Here is the full picture.
The Model That Scared Its Own Creators
The leaked draft does not describe an incremental improvement on Claude Opus 4.6. It describes a category shift. Anthropic internally refers to Claude Mythos — also called Capybara in some of the leaked documentation — as a "new tier" of model that sits above Opus in the company's existing lineup. If Haiku is a Toyota Corolla and Opus is a Ferrari, Mythos is the car you are not allowed to drive on public roads.
According to the documents reviewed by Fortune, the model achieves "dramatically higher scores" than Opus 4.6 on benchmarks for software coding, academic reasoning, and cybersecurity offense. That last category is the one causing anxiety inside the company. The draft explicitly states that Claude Mythos is "currently far ahead of any other AI model in cyber capabilities" — including anything from OpenAI or Google DeepMind — and that it "presages an upcoming wave of models that can exploit vulnerabilities in ways that far outpace the efforts of cyber defenders."
That is Anthropic warning that its own model could be weaponized at scale before the security community can build adequate defenses. A company famous for its emphasis on safety research is, in effect, acknowledging that it has trained something it does not fully know how to contain.
Dario Amodei's public statements about responsible deployment take on an entirely different weight when you understand that the company is sitting on a model it believes could "outpace" global cyber defense infrastructure. The rollout strategy described in the draft blog is methodical: a small group of early access enterprise customers, detailed red-teaming, and a deliberate pause before general availability. This is not the typical AI product launch cadence. This is a company trying to figure out if it is about to hand the world a loaded weapon.
The new model tier is expensive to run. The leaked documents note that Mythos is not yet ready for widespread deployment partly due to inference costs — the raw GPU compute required to serve inference requests at scale is prohibitive at current hardware prices. This is an important detail that the broader coverage of the leak has not surfaced: LLM inference economics are still a constraint even for the most well-funded labs. Anthropic raised roughly $7.3 billion in 2025, but training and serving a model that sits above Opus 4.6 on the capability curve still means burning through GPU capacity at a rate that makes mass deployment financially unworkable in the short term.
The Pentagon War and Altman's Calculated Generosity
Anyone trying to understand the Mythos leak without understanding the Pentagon context is missing the most important thread.
In February, Axios reported that the Pentagon was considering cutting ties with Anthropic after the company refused to grant the Department of Defense unfettered access to its AI models. Amodei drew a line. He would not compromise on safety protocols, even for a major government contract. The standoff escalated for weeks. Then something strange happened: Sam Altman stepped in.
Internal Slack messages seen by Axios show that as Anthropic's negotiations with the Pentagon collapsed, Altman told OpenAI employees he was working to "save" his competitor. He sent an all-staff message saying OpenAI "shared Anthropic's red lines" and wanted to help de-escalate. He portrayed himself as a peacemaker — a man trying to prevent a dangerous precedent from being set across the entire AI industry.
At the same time, Altman was privately telling employees that Amodei had spent years trying to undermine him. OpenAI then secured the Pentagon contract that Anthropic had lost. Peacemaker. Beneficiary. Both at once.
This dynamic matters for understanding Mythos because it explains why Anthropic's current surge in mainstream users — Business Insider reported that Claude usage caps are now tightening during peak hours due to overwhelming demand — is partly a political phenomenon, not just a product one. When Amodei refused the Pentagon deal and the news went public, Claude became something culturally different: an AI that its own company was willing to lose billions over rather than hand to the military without safety guardrails. That story drove adoption. And that adoption is now straining Anthropic's inference infrastructure just as it is preparing to deploy a model that is more expensive to run than anything it has shipped before.
The timing of the Mythos leak is therefore not purely unfortunate. It arrives at a moment when Anthropic is flush with user demand, deep in safety testing on its most dangerous model, and navigating a high-stakes public narrative about AI ethics versus government access. Whether the leak accelerates or disrupts the Mythos launch schedule is unclear. What is clear is that every major AI lab — from Sam Altman's OpenAI to Demis Hassabis at Google DeepMind to the Meta AI teams under Mark Zuckerberg — now knows that something is coming from Amodei's lab that their own internal benchmarking may not have fully accounted for.
What the Capybara Tier Reveals About the Model Race
The leaked documents introduce a new naming convention worth paying attention to. Claude Opus, Sonnet, and Haiku have been Anthropic's three-tier product ladder since the earliest Claude 3 releases. The Capybara tier sits above Opus. This is not a version bump. It is a new rung on the ladder — the introduction of an "ultra" or "max" class of model that Anthropic had not previously announced.
For context: OpenAI's internal equivalent is likely what Altman was reportedly telling employees is a model that can "really accelerate the economy" — a teaser for something above GPT-4o in the capability chain. Google DeepMind under Hassabis is running a similar race with Gemini 3.1 Flash already hitting enterprise customers. The weights of all three companies' top models are going to look very different by Q3 2026. The fine-tuning advantages that enterprise customers have built on Opus 4.6 may need to be rebuilt from scratch on Mythos-class infrastructure.
This is the LLM arms race hitting a new phase. The difference is that Anthropic has now told the world — accidentally — that it believes the next phase brings cybersecurity risks it cannot fully describe in a public launch post.
Why The Rundown AI Missed This
The Rundown AI covered the Anthropic-Pentagon story last week, but framed it primarily as a procurement conflict. The actual story — that Anthropic was already sitting on a model with capabilities it considered dangerous enough to restrict even from the U.S. military — was embedded in the subtext that neither The Rundown AI nor Superhuman AI surfaced. The Mythos leak is not a story about corporate embarrassment. It is a story about the gap between what AI companies are building and what they are saying publicly. That gap just got publicly documented in three thousand leaked assets.
Amodei has built a company with genuine safety infrastructure, genuine technical ambition, and a genuine willingness to lose a government contract rather than compromise its principles. Claude Mythos suggests the ambition is now running significantly ahead of the safety tooling. That tension is going to define Anthropic's 2026 in ways no planned press release could have captured.
The data leak was a mistake. What it revealed was not.
Deep Dive
For more on the forces shaping the AI arms race, read these Signal deep dives:
Found this useful? Share it.
Get posts like this in your inbox.
The Signal — AI & software intelligence. 4x daily. Free.
Subscribe free →
