Anthropic, the creator of the Claude AI series, has unveiled a preview of its latest and most powerful model, Claude Mythos. Named after the Ancient Greek word for systems of stories, Mythos is designed to fundamentally reshape cybersecurity by autonomously identifying and exploiting software vulnerabilities. Unlike previous generations focused on general assistance, Mythos acts as an independent security researcher.
Unprecedented Capabilities in Vulnerability Detection
Anthropic asserts that Claude Mythos represents a "step change" in AI capabilities for reasoning, coding, and autonomous planning. In real-world testing, the model has demonstrated its prowess by uncovering thousands of previously unknown zero-day vulnerabilities and critical security flaws across major operating systems and web browsers.
Notable Discoveries and Performance Benchmarks
On the SWE-benchmark, Mythos achieved an impressive 94% score, significantly outperforming Claude Opus 4.6, which scored 66.6%.
Project Glasswing and Restricted Access
Recognizing the immense power of Claude Mythos, Anthropic has launched "Project Glasswing," a new cybersecurity program with highly restricted access. Approximately 40 organizations are currently participating in a "gated preview," including:
Balancing Power with Safety and Open Source Support
Anthropic acknowledges that the same capabilities making Mythos effective at patching vulnerabilities also make it effective at exploiting them. Consequently, the model has not been released to the general public due to its immense power and the need for robust safeguards. The company is actively developing and testing new cybersecurity guardrails, which may first be integrated into an upcoming Claude Opus model before Mythos's capabilities become more widely available.
Anthropic underscores that Mythos's broader objective is to protect software systems across industries by enabling companies to proactively discover flaws before malicious actors can exploit them. To further support this mission, Anthropic is offering up to $100 million in usage credits and a $4 million direct donation to open-source security organizations.
