AISI backs Anthropic's Mythos: only model to finish a 32-step corporate network attack sim
dbreunig · x · 2026-10-22
Anthropic's unreleased security-focused model Mythos drew skepticism, but the first third-party analysis from the AI Security Institute largely supports the claims: "a step up over previous frontier models." The key test, "The Last Ones," is a 32-step corporate network attack simulation AISI estimates would take humans 20 hours. Across 10 runs each, Mythos was the only model to complete it — succeeding 3 of 10 times, versus Opus 4.6 and GPT-5.4. The author frames an emerging security economy: defenders must spend more tokens finding exploits than attackers spend exploiting, making cybersecurity a proof-of-work contest.
More from Models
- FOSS stack with DeepSeek beats paid ChatGPT plan in 2-month agent reliability test — layer4down · 2026-10-22
- Qwen3.8 27B builds Mont-Saint-Michel voxel diorama in one hour from a single prompt — Teknium · 2026-10-22
- Claude Opus 5.5 autonomously designs, edits and launches a $34 T-shirt end to end — zealcaiden · 2026-10-22
- Wikipedia goes down over two PHP flaws as OpenAI and humans spar over who found them — bindureddy · 2026-10-22
- Unsloth ships day-0 Dynamic GGUFs; local model nails tool calls, builds Flappy Bird in one shot — danielhanchen · 2026-10-22
- Reflection ships open Beam customization recipes spanning LoRA, SFT and GRPO RL — HowDevelop · 2026-10-22