Anthropic confirmed on April 20, 2026, that it is maintaining strict limitations on its most advanced artificial intelligence model, Claude Mythos Preview, due to the system's unprecedented capabilities in autonomous cybersecurity operations. The model, which the company has described as a step change in AI performance, is currently restricted to a private consortium of approximately 40 organizations under an initiative titled Project Glasswing. This restricted rollout marks a shift in the industry, where deployment constraints are being driven by national security concerns rather than commercial readiness.
The decision to withhold Mythos from general public release follows internal and third-party testing that identified the model as a potential offensive cyberweapon. According to technical documentation released by Anthropic, Mythos Preview achieved an 87.3% accuracy rate on the SWE-bench Pro coding benchmark, significantly outperforming previous frontier models like Claude Opus 4.6 and competitors such as GPT-5.4. Researchers reported that the model can independently scan millions of lines of source code to identify, and in some cases chain together, high-severity vulnerabilities with minimal human supervision. In one instance, the model reportedly discovered a 27-year-old vulnerability in legacy operating system code that had remained undetected by human engineers.
Reports surfaced on April 20 indicating that the U.S. National Security Agency (NSA) is among the entities utilizing the model for defensive testing. This development comes despite a March 2026 designation by the Department of Defense labeling Anthropic as a supply chain risk following the company's refusal to remove certain safety protocols that prohibit the use of its technology for autonomous weaponry. Anthropic CEO Dario Amodei met with White House Chief of Staff Susie Wiles and Treasury Secretary Scott Bessent on April 17 to discuss the strategic deployment of Mythos within critical infrastructure sectors, a meeting described by officials on Monday as productive and constructive.
The Project Glasswing consortium includes a range of systemic financial institutions and technology leaders, such as JPMorgan Chase, Goldman Sachs, Citigroup, Amazon, Apple, and Microsoft. These organizations are utilizing the model to conduct red-team exercises and preemptively patch zero-day vulnerabilities within their own networks. Anthropic stated that the model has already identified thousands of previously unknown flaws across every major operating system and web browser currently on the market.
Global regulators have responded to the restricted rollout with increased vigilance. On April 20, the Australian Securities and Investments Commission (ASIC) confirmed it is monitoring the implications of Mythos for financial market stability, joining other Asia-Pacific authorities in tracking the technology's development. While Anthropic released a less capable version, Claude Opus 4.7, for general use on April 16 with integrated cyber-safeguards, the company maintains that Mythos-class models will remain restricted until more robust security frameworks are established. Anthropic has emphasized that the current deployment strategy is designed to ensure that defensive capabilities outpace the potential for malicious exploitation by state and non-state actors.