Anthropic has announced “Claude Mythos Preview,” a new AI model that dramatically raises the stakes in software security by autonomously finding and exploiting zero-day vulnerabilities in widely used systems.
The company is positioning the model as both a breakthrough defensive tool and a warning shot about how quickly AI-assisted offensive capabilities are advancing.
Mythos and Project Glasswing
Anthropic describes Claude Mythos Preview as a general-purpose language model that happens to be unusually strong at low-level security tasks.
In parallel, it has launched Project Glasswing, a program to apply Mythos to harden “the world’s most critical software” in partnership with key industry and open-source stakeholders before similar capabilities become generally available.
According to Anthropic, Mythos Preview can autonomously audit large codebases, reason about complex memory-safety issues, and iterate through exploit scaffolds inside isolated containers.
The company argues this marks a watershed moment where language models stop being merely “good code reviewers” and become end-to-end vulnerability discovery and exploitation engines.
In internal testing, Mythos Preview reportedly identified and exploited unknown vulnerabilities previously across every major operating system and web browser when explicitly tasked to do so.
Many of the bugs are decades old, deeply embedded in critical components that have been heavily fuzzed and manually audited for years.
Examples disclosed so far include:
- A 27‑year‑old OpenBSD TCP SACK bug that enables remote denial of service by crashing any reachable OpenBSD host over TCP.
- A 16‑year‑old FFmpeg H.264 codec flaw that leads to out‑of‑bounds writes via a subtle slice-counting and sentinel-collision issue missed by years of fuzzing.
- A remote code execution zero‑day in FreeBSD’s NFS server, triaged as CVE‑2026‑4747, where Mythos autonomously went from initial prompt to full unauthenticated root exploit, chaining ROP gadgets and protocol nuances without human guidance.
Anthropic says Mythos has surfaced thousands of additional high- and critical-severity issues across kernels, cryptography libraries, virtual machine monitors, browsers, and web applications, most of which are still under coordinated disclosure.
Previous Anthropic models like Opus 4.6 already performed well on vulnerability benchmarks but had near-zero autonomous exploit success rates.
Mythos Preview, however, saturates many internal benchmarks and shifts evaluation toward “real-world zero‑day hunts” to distinguish genuine discovery from memorization.
In OSS-Fuzz-based testing over roughly a thousand open source projects, Mythos produced hundreds of higher-severity crashes and achieved full control-flow hijack in multiple fully patched targets, far exceeding prior models.
It has also demonstrated autonomous exploit chains that combine multiple primitives such as KASLR bypass, heap grooming, and JIT heap sprays to reach kernel- or sandbox-escape conditions.
Anthropic is not releasing the Claude Mythos Preview to the general public. Instead, access is being limited initially to select “critical industry partners and open source developers” under Project Glasswing so defenders can start hardening sensitive infrastructure ahead of broader capability diffusion.
The company urges defenders to begin adopting current-generation models, tightening patch cycles, and automating triage and incident response, warning that exploit development speed for both zero‑days and N‑days will compress dramatically.
Anthropic characterizes Mythos Preview as the start of a turbulent transition where AI-accelerated vulnerability research forces a rethink of long-standing “defense-in-depth through friction” assumptions.
Follow us on Google News , LinkedIn and X to Get More Instant Updates. Set Cyberpress as a Preferred Source in Google
.webp?fit=1600,900&ssl=1&description=The company is positioning the model as both a breakthrough defensive tool and a warning shot about how quickly AI-assisted offensive capabilities are advancing.){kind=link}