Anthropic Expands Public Access to Claude Mythos AI Model
Anthropic plans to expand access to its advanced AI model, Claude Mythos, through Project Glasswing, initially to U.S. and allied governments and select partners, before a broader release in 6-12 months. The model has already uncovered over 10,000 high- or critical-severity vulnerabilities in critical software, prompting concerns about its potential misuse by threat actors.
Anthropic, the AI research company, announced plans to expand access to its powerful Claude Mythos model, initially restricted due to its ability to uncover and exploit software vulnerabilities. The company stated it will work with critical partners, including U.S. and allied governments, to broaden participation in Project Glasswing, a program currently involving around 50 organizations like Cisco, Oracle, and Microsoft. Mythos has already identified over 10,000 high- or critical-severity flaws in globally critical software, with 6,202 found in open-source projects. A third-party review confirmed 91% of 1,752 examined vulnerabilities were legitimate, with two-thirds correctly rated. Mythos stands out for its capability to generate working exploit chains, sometimes combining minor flaws into severe threats, prompting Anthropic’s cautious approach. The company warned that attackers will soon have access to similar tools, predicting Mythos-level models will be widely available within 6-12 months. While details on discovered vulnerabilities remain limited to avoid endangering users, Britain’s AI Security Institute confirmed Mythos could autonomously execute multi-stage attacks in controlled tests, though its effectiveness against well-defended systems remains unproven. Anthropic emphasized the urgency of safeguarding against potential misuse while accelerating vulnerability discovery. The company’s partners have prioritized patching flaws before malicious actors exploit them, though the long-term risks of advanced AI-driven cyber threats remain unclear. As frontier AI models evolve, new benchmarks are being developed to measure their exploit capabilities, signaling a shift in cybersecurity dynamics. The expansion of Project Glasswing reflects Anthropic’s strategy to balance accessibility with risk mitigation, aiming to fortify global software security before adversaries gain equivalent tools. Meanwhile, the AI Security Institute’s findings underscore the growing sophistication of AI in cyber operations, raising questions about future defensive strategies in an increasingly automated threat landscape.
This content was automatically generated and/or translated by AI. It may contain inaccuracies. Please refer to the original sources for verification.