Claude Mythos Preview

Discovered cryptographic weaknesses in HAWK and reduced-round AES after 60 hours of continuous operation

cost:

$100,000

duration:

60 hours

Anthropic restored Mythos access with a security deal after the vulnerability spike.

Built 8 working exploits from Firefox and Windows kernel patches in ~12 hours; first exploit ready 18 days before patched Firefox 148 shipped

METR found Claude Mythos Preview could work 16+ hours autonomously

autonomous work duration:

16+ hours

First AI model to clear all UK AISI cyberattack simulations

Early snapshot achieves more than 2x time horizon of next best model on METR benchmark

Claude Mythos Preview fully solved TLO enterprise network simulation in 3 of 10 attempts

simulation:

TLO

success rate:

3/10

Claude Mythos Preview scored 68.6% on AISI expert CTF tasks

score:

68.6%

task type:

expert CTF

First AI model to complete full AISI cybersecurity evaluation

Leaked benchmark results suggest it 'destroys every other model' including GPT-5, Claude 4 Opus, and Gemini Ultra 2.0

Achieved 73% success rate on expert-level CTF challenges and completed full 32-step network attack simulation

ctf success rate:

73%

network attack success:

3 of 10 attempts

Evaluated by UK AI Safety Institute for autonomous cyber attack capabilities.

evaluator:

UK AI Safety Institute

test type:

capture-the-flag and network simulation

First AI model documented to autonomously complete a full, multi-step cyber attack simulation in UK safety tests.

success rate:

73%

simulation success:

3 of 10 attempts

steps:

32-step network takeover

Scored 83.1% on the CyberGym benchmark for vulnerability discovery.

score:

83.1%

benchmark:

CyberGym

Anthropic launched Claude Mythos Preview, capable of autonomous cybersecurity vulnerability discovery and exploitation.

Anthropic released Claude Mythos Preview model, used by Mozilla for security bug triage and patching.