Anthropic to Permanently Withhold the Mythos AI Model from Public Release

According to monitoring by Beating, Anthropic has reaffirmed in the latest progress update of its software security collaboration initiative, Project Glasswing, that despite internal testing demonstrating that its next-generation safety-aligned foundation model, Claude Mythos Preview, possesses exceptional capabilities in vulnerability discovery and cyber defense, the company will indefinitely prohibit public and commercial release of Mythos-level models until more effective safety alignment mechanisms are developed. In an AI industry landscape marked by fierce competition to release larger models and lower API pricing, this decision not to publicly release the model is notably restrained. Based on joint testing feedback, while Mythos Preview significantly reduces the cost for security professionals to identify vulnerabilities, unrestricted public release would drastically lower the cost of developing malware and launching cyberattacks, posing catastrophic weaponization risks to global internet infrastructure—both physical and digital. Over the past month, in joint offensive and defensive tests, Anthropic, alongside approximately 50 partners including Microsoft, Oracle, Cloudflare, and Mozilla, used Mythos Preview to identify over 10,000 high- or critical-severity zero-day vulnerabilities across critical global infrastructure. The UK AI Safety Institute (UK AISI) assessed that Mythos Preview is the first model globally to fully breach all of its simulated cyber defense ranges. Although vulnerability discovery efficiency has increased more than tenfold—such as Mozilla leveraging Mythos Preview to discover and patch 271 vulnerabilities in Firefox 150, more than ten times the number found using Claude Opus 4.6—the security ecosystem remains in an unprecedented state of “patch overload.”