Anthropic used Claude to prove 123 OSS libraries crash-free, with 180M lines of Rocq proofs
inductionheads · x · 2026-10-23
Anthropic's Frontier Red Team blog details work with Theorem Labs: in a few weeks, Claude patched bugs and proved crash-freedom on any input (up to 64MB) for 239 entry points across 123 OSS-Fuzz libraries.
- Proofs operate on compiled RISC-V binaries: 180M lines of Rocq total, 600K lines per proof, covering binaries with a median 18,000 instructions
- Method: translate RISC-V assembly into a mathematical model via Theorem's instruction and Linux process semantics; assumes single-threaded static binaries with fixed memory budgets
- Proofs are reusable: agents sharing a workspace copied each other's proofs, cutting token use by 17%; 324 files reused across 36 proofs
- Sparks debate that formal verification flips the attacker's long-held advantage
More from Safety
- Oxford paper draws TCP/IP lessons for securing and governing MCP and A2A agent protocols — sethlazar · 2026-10-23
- Cloudflare CEO: AI-uncovered vulnerabilities could mean weekly security crises for years — kimmonismus · 2026-10-23
- Cloudflare CEO warns AI is unearthing hidden vulnerabilities, years of weekly crises ahead — kimmonismus · 2026-10-23
- AI Now's Heidy Khlaaf to keynote Black Hat Europe on why AI Security ≠ AI Safety — AINowInstitute · 2026-10-23
- Morgan Stanley is blocking AI agents — a 6-point playbook for retailers before their checkout is next — sanjaykalra · 2026-10-23
- AI found 3% of all critical software bugs ever in one month, on pace for 100K this year — adivinemessenger · 2026-10-23