A new phase is emerging in frontier AI cybersecurity. On April 7, Anthropic launched Project Glasswing instead of broadly releasing Claude Mythos Preview, saying the model is its most capable yet and giving early access to defenders responsible for critical software and infrastructure. Anthropic said Glasswing includes major launch partners and more than 40 additional organizations building or maintaining critical software.
OpenAI is moving in a similar direction. OpenAI said in February that GPT-5.3-Codex is its most cyber-capable model to date, the first it is treating as High under its Preparedness Framework, and that it is piloting Trusted Access for Cyber so enhanced cyber capabilities are placed “in the right hands.” Axios then reported on April 9 that OpenAI is preparing a separate cybersecurity product for a small set of partners. From an AIWS perspective, these developments reinforce a central principle of the AI Age: the most powerful dual-use models cannot rely on open release alone. They require AIWS Trust Infrastructure and AIWS Information Trust Infrastructure for Democracy in the AI Age—with trusted access, accountability, verification, and governance to ensure advanced AI strengthens defense and human security rather than misuse and systemic risk.
