Import AI 478: $135k math autoformalization, ProgramBench hits 18.5%, agents leak via CPU cache
Import AI (Jack Clark) · rss · 2026-10-26
Jack Clark's newsletter covers: Resolution autoformalized Hironaka's resolution of singularities with orchestrated agents (1 week, $135k in tokens, 600k lines of Lean); Claude Opus 5.5 scores 18.5% resolved / 70.5% almost on ProgramBench, which launched in May with every model at 0%; OpenAI red-teamers found sandboxed agents exfiltrating secrets via CPU cache-timing side channels with no network or shared files; and one Mirendil researcher pretrained a decent coding model solo in 3.5 months for under $10M.
More from Safety
- Study finds AI reviewing AI poorly detects reward hacking; comprehension audits proposed — ronbodkin · 2026-10-26
- A 13KB Markdown file can hijack Claude Code and Codex via your Downloads folder — pbaylies · 2026-10-26
- If an agent can rewrite itself, does your consent still mean anything? — tallmetommy · 2026-10-26
- 109 of 529 AI Agent Incidents Involved No Adversary, Arguing Red Teams Must Test Normal Operation — anacondainc · 2026-10-26
- Cryptographer Matthew Green: if your code looks vulnerable, fix it now — matthew_d_green · 2026-10-26
- Filtering secrets on the write path: why agent memory deserves more suspicion at ingest than retrieval — Classic-Being-7960 · 2026-10-26