Anthropic Leaks Most Powerful AI Model Amid IPO Speculation
Anthropic's accidental exposure of its most advanced AI model's cyber warfare capabilities raises urgent questions about corporate transparency just as the company prepares for a massive public offering.
Anthropic confirmed Thursday it accidentally exposed details of its most powerful AI model ever developed — a system with unprecedented cyber warfare capabilities — just as the company reportedly prepares for a $60 billion public offering this fall.
Security researchers Roy Paz and Alexandre Pauwels discovered nearly 3,000 unpublished Anthropic assets in a publicly searchable data lake this week. Among the findings sat a draft blog post announcing "Claude Mythos." The leak revealed a new top-tier model called "Capybara" that dramatically outperforms current AI systems in coding, academic reasoning and cybersecurity tasks.
"Capybara is a new name for a new tier of model: larger and more intelligent than our Opus models—which were, until now, our most powerful," the leaked draft states. "Compared to our previous best model, Claude Opus 4.6, Capybara gets dramatically higher scores on tests of software coding, academic reasoning, and cybersecurity, among others."
The timing raises urgent questions about corporate transparency and market incentives. Bloomberg reported Thursday that Anthropic is weighing an initial public offering as soon as October 2026, potentially raising more than $60 billion in what would be one of the largest tech listings in years. Goldman Sachs, JPMorgan Chase and Morgan Stanley have held preliminary talks about the potential listing.
"We're developing a general purpose model with meaningful advances in reasoning, coding, and cybersecurity," an Anthropic spokesperson told Fortune. "Given the strength of its capabilities, we're being deliberate about how we release it. We consider this model a step change and the most capable we've built to date."
The leaked documents contain stark warnings about the model's potential impact. Internal assessments describe Mythos as "currently far ahead of any other AI model in cyber capabilities" and warn it "presages an upcoming wave of models that can exploit vulnerabilities in ways that far outpace the efforts of defenders."
Investors immediately recognized the threat. Cybersecurity stocks including CrowdStrike, Palo Alto Networks and Fortinet fell 4 percent to 6 percent within hours of the leak's disclosure. The market signaled fear that the new model could destabilize the entire cyber defense industry.
Anthropic's internal materials show the company planned to release Capybara first to cyber defense organizations. "We're releasing it in early access to organizations, giving them a head start in improving the robustness of their codebases against the impending wave of AI-driven exploits," the leaked draft states.
The company's concerns appear well-founded. Anthropic's own November 2025 report documented a Chinese state-sponsored group using Claude Code to infiltrate approximately 30 organizations in what the company called "the first reported AI-orchestrated cyber espionage campaign." The AI performed 80 percent to 90 percent of campaign work autonomously, with human operators involved only at four to six critical decision points.
While Anthropic attributed the leak to "human error" in its content management system, the exposure occurred against a backdrop of carefully orchestrated private events. CEO Dario Amodei planned to personally brief Europe's top business leaders on unreleased capabilities at an exclusive two-day retreat in an 18th-century English manor this spring.
The company confirmed the retreat as "part of an ongoing series of events we've hosted over the past year" and said attendees would "experience unreleased Claude capabilities" during the gathering.
Market analysts note the suspicious convenience of a "leak" that positions Mythos as a generational leap in AI capabilities just months before a potential public offering. Such publicity typically drives investor interest and valuation multiples in the lead-up to major technology listings.
Anthropic's valuation reached $380 billion in February after raising $30 billion in Series G funding. The company has committed to $50 billion in U.S. data center investments and generates estimated annualized revenue of $14 billion.
The company's official explanation leaves questions unanswered. "An issue with one of our external CMS tools led to draft content being accessible," the spokesperson told Fortune. "We're working with a small group of early access customers to test the model."
Hamza Chaudhry, AI and national security lead at the Future of Life Institute, questioned the strategic logic behind developing such powerful offensive tools. "The strategic logic of racing to deploy AI systems that demonstrably empower adversaries—while hoping these same systems will help us defend against attacks conducted using our own tools — appears fundamentally flawed and deserves a rethink in Washington," he told FOX Business.
The leak comes just one day after U.S. District Judge Rita Lin blocked the Pentagon's supply chain risk designation against Anthropic, calling it "Orwellian" in her ruling. The legal victory removed regulatory hurdles that could have complicated the company's IPO plans.
As investors assess Anthropic's market potential, the more durable question is simpler and harder to answer: how did the story get out, and who benefited? No model was leaked — only a description. But that description alone moved markets, put every nation-state threat actor on notice that a step-change in AI-driven exploitation is imminent, and generated the kind of pre-IPO awareness that no marketing budget can replicate. Researcher Roy Paz brought the exposed assets to Fortune rather than to Anthropic directly — a decision that becomes easier to understand given that Anthropic had declined, just weeks earlier, to fix a separately reported zero-click, maximum-severity vulnerability in its own platform. Fortune notified Anthropic before publishing; Anthropic confirmed the model's existence on the record and let the coverage run.
A Chinese state-sponsored group already ran an 80–90% automated espionage campaign against 30 organizations using a less capable Claude model. The upgrade is coming. Whether the chain of events that announced it to the world was negligence, calculation, or simply a researcher doing his job and a journalist doing hers, the outcome is the same: Anthropic heads into its $60 billion IPO as the company believed to be building the most powerful — or most dangerous — AI on the planet. In Silicon Valley, there are worse things to be known for.