Anthropic Restricts Access to ‘Dangerous’ Mythos AI Model via Project Glasswing
Anthropic has announced a restricted rollout of its most advanced artificial intelligence model to date, Claude Mythos Preview, citing significant concerns over its potential to be weaponized for cyberattacks. In an unprecedented move for a frontier model release, the company is limiting access to a select group of partners to ensure the technology is used exclusively for defensive security purposes.
The model, which Anthropic describes as a striking leap in capability over its previous frontier model, Claude Opus 4.6, possesses advanced agentic coding and reasoning skills. According to company reports, these capabilities craft the model exceptionally proficient at identifying software weaknesses and security flaws—a double-edged sword that could either bolster global defenses or provide disappointing actors with a powerful tool for exploitation.
To manage this risk, Anthropic has established Project Glasswing, a cybersecurity initiative that grants a curated group of Substantial Tech and security firms access to the preview version of Mythos. The partnership includes industry giants such as Microsoft, Amazon Web Services, Apple, Google, and Nvidia, as well as specialized cybersecurity firms like CrowdStrike and Palo Alto Networks. Other critical infrastructure participants include JPMorganChase, Cisco, and Broadcom, alongside approximately 40 other organizations.
This strategic limitation underscores the growing tension between AI innovation and global security. The decision highlights the industry’s struggle to balance the deployment of high-capability models with the necessity of preventing large-scale, AI-driven cyber threats.
The existence of the Mythos model first became public in March, following a report by Fortune after descriptions of the model were found in a public data cache. A leaked draft blog post from the company warned that Mythos was “currently far ahead of any other AI model in cyber capabilities” and could potentially outpace the efforts of human defenders to secure vulnerabilities.
“There was a lot of internal deliberation,” Dianne Penn, Anthropic’s head of research product management, stated in an interview. “We really do view this as a first step for giving a lot of cyber defenders a head start on a topic that will be increasingly important.”
While Anthropic has no immediate plans to make the Mythos Preview generally available to the public, the company stated its long-term goal is to safely deploy Mythos-class models at scale once more robust safeguards are in place. For now, the focus remains on using the model’s ability to discover security gaps to harden critical software and open-source code across the global digital ecosystem.
The move has already sent ripples through the market; reports indicate that cybersecurity stocks experienced volatility following the initial disclosure of the model’s capabilities, reflecting investor anxiety over the potential for AI to disrupt traditional security paradigms.
As the industry moves forward, the rollout of Mythos serves as a critical test case for how AI labs will manage “too dangerous” technologies in an era of accelerating intelligence.
