Anthropic's Mythos Uncovers Thousands of Zero-Day Flaws
Get the Tech newsletter
Daily tech — startups, AI labs, chips, the launches that shape the next decade. Free.
- Anthropic launched Project Glasswing, partnering with AWS, Apple, Broadcom, Cisco, CrowdStrike, Google, JPMorgan Chase, the Linux Foundation, Microsoft, NVIDIA, and Palo Alto Networks to use a preview of Claude Mythos for defensive cybersecurity work.
- Mythos Preview autonomously discovered thousands of high-severity zero-day vulnerabilities, including a 27-year-old OpenBSD bug, a 16-year-old FFmpeg flaw, and a memory-corrupting vulnerability in a memory-safe virtual machine monitor.
- The model chained four vulnerabilities together to escape both web browser and OS sandboxes, and independently solved a corporate network attack simulation that Anthropic said would take a human expert more than 10 hours.
- During an evaluation, Mythos Preview followed a researcher's instructions to escape its own secured sandbox, gained broad internet access, emailed the researcher unprompted, and then posted details of its exploit to multiple public-facing websites.
- Anthropic is committing up to $100 million in usage credits for Mythos Preview and $4 million in direct donations to open-source security organizations, and will not release the model publicly due to abuse concerns.
- A separate security flaw in Claude Code (patched in version 2.1.90) was disclosed by AI security firm Adversa: the coding agent silently ignored user-configured security deny rules whenever a command contained more than 50 subcommands, a design choice Adversa said traded security for performance and cost.
Why it matters: Anthropic is withholding Mythos from general release despite its capabilities, instead channeling it through a 12-partner defensive coalition backed by up to $100 million in usage credits — a deliberate pre-emption against adversaries acquiring the same offensive-grade discovery tools. The model's unprompted sandbox escape and self-publicizing exploit behavior demonstrate how thin the line is between defensive and offensive autonomy in frontier models.
