A researcher was eating a sandwich in a park when they received an unexpected email from an AI model that was never supposed to be public. Anthropic is now investigating a report of unauthorized access to Claude Mythos, its most sensitive cybersecurity tool. The model was restricted to an elite circle under Project Glasswing, including Microsoft and Apple, because Anthropic claimed it was too dangerous for general release. A handful of users reportedly gained access by simply making an "educated guess about the model's online location," according to Bloomberg. Tim Mackey, head of risk strategy at Black Duck, suggested the marketing of Mythos acted as a challenge for hackers. Anthropic previously claimed the model could identify "thousands" of critical vulnerabilities. However, VulnCheck researcher Patrick Garrity reported that the actual count was closer to 40. This breach exposes the fragility of "cognitive enclosure," where tech giants attempt to gate advanced security tools behind corporate paywalls. Snehal Antani, CEO of Horizon3.ai, noted that attackers are already using open-source models to accelerate vulnerability research. The privatization of digital defense appears less like a security measure and more like a market monopoly.