Assessing Claude Mythos Preview’s cybersecurity capabilities
Anthropic’s new Claude Mythos Preview model can autonomously find and exploit zero-day and N-day vulnerabilities across major OSes and browsers. In testing, it produced sophisticated exploits—from JIT heap sprays to multi-packet ROP chains—and outperformed prior models by a wide margin. Project Glasswing will share these capabilities with select partners to shore up defenses before wider release.
- Claude Mythos Preview autonomously found and exploited a 27-year-old OpenBSD bug and chained four browser flaws into a JIT heap spray exploit, plus RCE on FreeBSD's NFS server via a 20-gadget ROP chain split across packets
- On Firefox JS engine trials, it produced 181 working shell exploits versus Opus 4.6's 2 successes in hundreds of attempts
- On OSS-Fuzz benchmarks (~7,000 entry points), it achieved full control-flow hijack (tier 5) on ten patched targets, where prior models never exceeded a single tier 3 crash
- These exploitation abilities emerged as a side effect of general code reasoning improvements, not targeted exploit training, prompting Anthropic to share the model early with defenders via Project Glasswing
On April 7, Anthropic unveiled Claude Mythos Preview, a general-purpose language model with a clear edge in computer security tasks. Under Project Glasswing, they’re using Mythos Preview to hunt zero-day flaws in critical software and to push the industry toward new defensive practices. In internal tests against major operating systems and web browsers, the model autonomously found and exploited vulnerabilities decades old—one 27-year-old bug in OpenBSD among them—and built multi-stage exploits, from JIT heap sprays chaining four browser flaws to remote code execution on FreeBSD’s NFS server via a 20-gadget ROP chain split over several packets.
Anthropic compared Mythos Preview’s performance to earlier models like Opus 4.6. While Opus managed successful autonomous exploits only twice out of hundreds of Firefox JavaScript engine trials, Mythos Preview produced 181 working shell exploits and achieved partial register control in 29 more. On OSS-Fuzz benchmarks covering about 7,000 entry points, previous models hit low-severity crashes (tiers 1–2) a few hundred times and saw a single tier 3 crash. Mythos Preview repeated 595 tier 1–2 crashes, added several tier 3–4 crashes, and reached full control-flow hijack (tier 5) on ten patched targets.
These offensive skills emerged without explicit exploit training; they’re a byproduct of stronger code understanding, reasoning, and autonomy. Anthropic argues that just as fuzzers evolved from attacker tools into core defensive assets, advanced language models can bolster software security—provided defenders gain early access. To that end, Project Glasswing is distributing Mythos Preview to select industry partners and open-source developers, hoping to lock down the highest-risk systems before similar models become widely available.