Claude Mythos: Anthropic’s High-Risk AI Too Powerful for Public Release
Anthropic announced on Tuesday the launch of Claude Mythos Preview, an advanced artificial intelligence model designed to identify security flaws and weaknesses within software.
Access to the model is being restricted to a select group of companies under a new cybersecurity initiative called Project Glasswing. This group includes Microsoft, Amazon Web Services, Apple, Google, and Nvidia, alongside more than 40 other organizations responsible for building or maintaining critical software infrastructure, including CrowdStrike, Palo Alto Networks, Broadcom, Cisco, and JPMorganChase.
Project Glasswing and Defensive Security
Under the terms of Project Glasswing, participating firms will use the Mythos Preview for defensive security work, scanning both their own proprietary systems and open-source code. Anthropic stated that these organizations will share their findings with the wider industry to bolster defenses across critical global systems.
Dianne Penn, Anthropic’s head of research product management, described the initiative as a first step in providing cyber defenders with a head start on a topic that will become increasingly important.
Cybersecurity Risks and Model Capabilities
The decision to limit the rollout follows internal deliberations regarding the model’s potential for misuse. Anthropic has warned that Claude Mythos Preview possesses agentic coding and reasoning skills that could reshape the cybersecurity sector, but similarly pose unprecedented risks by increasing the likelihood of large-scale AI-driven cyberattacks.
In a draft blog post that was inadvertently made public last month, Anthropic noted that the model is currently far ahead of any other AI model in cyber capabilities. The company warned that the model presages a wave of AI that can exploit vulnerabilities in ways that far outpace the efforts of human defenders.
Disclosure and Technical Benchmarks
The existence of the Mythos model was first revealed in March after Fortune reported on descriptions discovered in a publicly accessible data cache. The documents described the model as the most powerful AI Anthropic had ever developed, a report that contributed to a decline in cybersecurity stocks.
According to the company’s system card, Claude Mythos Preview shows a striking leap in scores on multiple evaluation benchmarks compared to the previous frontier model, Claude Opus 4.6.
Anthropic has stated it does not plan to make the Mythos Preview generally available. The company intends to deploy Mythos-class models at scale only once new safeguards are in place.