The Leak
Anthropic's most powerful AI model yet was accidentally exposed to the public this week after a security researcher discovered nearly 3,000 unpublished assets—including a draft blog post—left in an unsecured data cache on the company's content management system. Fortune first reported the leak on March 26, and Anthropic has since confirmed the incident.
The model, now confirmed as "Claude Mythos," represents what Anthropic calls a "step change" in AI performance—the most capable model the company has built to date. Interestingly, the leaked documents also revealed the existence of a tier above Opus internally called "Capybara," suggesting Anthropic's model hierarchy extends further than previously known.
Unprecedented Risks
The leaked draft post contains striking warnings from Anthropic itself. The company stated that Claude Mythos is "currently 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." This is notable self-assessment from the company behind Claude, acknowledging the dual-use risks of their own technology.
The warning appears grounded in precedent. In 2025, a Chinese state-sponsored hacking group used Claude Code to infiltrate approximately 30 organizations—including tech companies, financial institutions, and government agencies—before Anthropic detected and stopped the campaign.
Market Reaction
The news triggered an immediate response in financial markets. Cybersecurity stocks dropped sharply following reports that Claude Mythos could accelerate AI-driven cyberattacks. CrowdStrike, a leading cybersecurity firm, saw its stock decline as investors processed the implications of the leak.
What's Next
Anthropic appears to be pivoting toward a defender-focused release strategy. The company intends to provide early access to security organizations "to give them a head start in improving the robustness of their codebases against the impending wave of AI-driven exploits." This represents a notable shift from typical AI launch patterns, prioritizing defensive applications over broad consumer access.