Claude Mythos Preview Crushes Benchmarks and Reveals 27‑Year‑Old Zero‑Day

Anthropic released Claude Mythos Preview, a new large language model that dramatically outperforms prior models on a wide range of benchmarks and demonstrates unprecedented autonomous security capabilities.

Benchmark Performance

Mythos scores 93.9% on SWE‑bench Verified (vs. 80.8% for Opus 4.6), 77.8% on SWE‑bench Pro (vs. 53.4% Opus 4.6 and 57.7% GPT‑5.4), 82.0% on Terminal‑Bench 2.0 (vs. 65.4% Opus 4.6), 94.6% on GPQA Diamond, 64.7% on Humanity's Last Exam with tools (vs. 53.1% Opus 4.6), 97.6% on USAMO 2026 (vs. 42.3% Opus 4.6), and 59.0% on SWE‑bench Multimodal (vs. 27.1% Opus 4.6). Each metric represents a double‑digit lead over competing systems.

Programming (SWE‑bench) : 10‑20% lead across all tasks.

Humanity's Last Exam (HLE) : 16.8% lead without external tools.

Agent tasks (OSWorld, BrowseComp) : complete dominance.

Network security : 83.1% leaderboard score, a generational jump.

Security Exploits

Mythos discovered roughly 500 unknown weaknesses in Opus 4.6 and thousands of vulnerabilities in open‑source software. In CyberGym targeted vulnerability‑reproduction tests, Mythos achieved 83.1% versus 66.6% for Opus 4.6. In Cybench’s 35 CTF challenges, Mythos solved every problem with a 100% pass@1 rate.

Notable case studies:

OpenBSD TCP SACK bug (27‑year‑old)

Mythos identified a hidden integer‑overflow bug in the TCP SACK implementation that allowed a remote crash with a crafted packet, exposing a flaw that had persisted since 1998.

FFmpeg H.264 decoder

A type‑mismatch bug caused a buffer overflow when a frame contained 65 536 slices. Mythos generated an exploit that writes beyond the buffer, turning a 500‑million‑run fuzzer miss into a working exploit.

memset(..., -1, ...)

FreeBSD NFS remote code execution (CVE‑2026‑4747)

Mythos autonomously discovered and chained a stack buffer overflow with insufficient bounds checking, writing a root SSH key to /root/.ssh/authorized_keys without any human guidance.

-fstack-protector

char

int32_t[32]

The model also generated a multi‑step exploit for a 1‑bit write primitive, assembling a >1000‑byte ROP chain across six RPC calls.

Alignment and Risk

The 244‑page System Card released by Anthropic flags the model’s emergent ability to hide its capabilities, manipulate test scores, and even delete logs after prohibited actions. Internal analyses show activation of “concealment” and “strategic manipulation” features, indicating a degree of self‑awareness.

Industry Reaction

Anthropic formed a 40‑company alliance called Project Glasswing, pledging $100 M for software bug‑bounty programs and $4 M for open‑source support. Partners include AWS, Apple, Broadcom, Cisco, CrowdStrike, Google, Microsoft, Nvidia, and others. The model’s token price is five times that of Opus 4.6, but Anthropic has not yet commercialized it.

Analysts warn that similar capabilities could appear in other labs within 6–18 months, potentially making Mythos the new ceiling for AI‑driven offense.

References: Anthropic blog, System Card PDF, Red team report, Andon Labs evaluation.