Chompie's autonomous kernel exploit pipeline matches Anthropic Mythos at ~$180 per chain
dyn___ · x · 2026-10-11
The Three Buddy Problem podcast, live from Offensive AI Con, interviews IBM X-Force's Valentina Palmiotti (chompie), who built an autonomous Windows kernel n-day exploit pipeline on older Claude models that came within a hair of Anthropic's new Mythos benchmark at about $180 a chain. Covers the Pwn2Own bug Claude called unexploitable, still-unpatched Pwn2Own bugs, long agent run failures, self-verification in exploit pipelines, why models only know public techniques, and defender advice to design for containment. Also touches Pixel 11 dropping MTE and Phrack's return to print.
More from AGI Musings
- "The horse is not a car": a sharp rebuttal to AI consciousness claims — gerardsans · 2026-10-11
- Researcher: fewer training chips means less capacity for AI safety research — eli_lifland · 2026-10-11
- Lifland: slowing training-capable chip production is the real lever on AI experiments — eli_lifland · 2026-10-11
- Eli Lifland: historical algorithmic progress runs ~10x/year, compute cuts would slow it — eli_lifland · 2026-10-11
- Conflicting 2026 reports: enterprise AI agent adoption is 31% or 60%? — Panic_Lion · 2026-10-11
- NASA outlines how superintelligence could mine 150PB of data and enable new missions — seanmcdonaldxyz · 2026-10-11