Brad Bettridge, a security professional working within a hub called Arena, has competed in over 1,000 challenges to find holes in the world’s most powerful AI models. His work is part of an 'army' of 15,000 hackers employed by Gray Swan to pressure-test systems like Anthropic’s Claude Mythos and OpenAI’s GPT-5. As these private labs identify thousands of 'zero-day' vulnerabilities, they are simultaneously enclosing that knowledge behind proprietary paywalls. This 'Cognitive Enclosure' was codified this week as Anthropic prepared the public rollout of Mythos through 'Project Glasswing,' a gated initiative that limits access to a small group of corporate and military elites, including the Pentagon. While these firms sell security as a luxury service, the public digital commons is rotting. The Verizon 2026 Data Breach Investigations Report revealed that threat actors are now using AI to scale attacks, highlighting 'VoidLink,' a malware framework assembled by an agent in just six days. Verizon’s research found that the median attacker now utilizes AI across 15 distinct techniques, even as corporate employees inadvertently leak source code through unauthorized AI usage. The hum of server racks and the blinking of fiber-optic cables in private data centers mark the boundary of this new digital perimeter. For the working class, this enclosure means the terminal loss of digital sovereignty; the state, crippled by data leaks like the recent CISA master key exposure, is now forced to lease its defense from the very monopolies that outpaced it. According to the New Zealand National Cyber Security Centre, the speed at which models like Mythos can detect flaws has caused a 'significant increase' in security incidents, forcing banks to rush to plug holes. This is the logic of the protection racket applied to the 21st century: the tech giants build the monster, then charge the public for a cage.