At about 06:01 Beijing time on October 23, 2026, the Anthropic account on X pointed to a Frontier Red Team research post dated October 22, 2026. Anthropic says that in a collaboration with Theorem, Claude formally proved security properties at the RISC-V assembly level for real open-source libraries from Google’s OSS-Fuzz. DigIntoAI has not tested the proofs or the pipeline.
Per Anthropic, Claude patched bugs it found and then proved there are no crashes on any input under 64 MB for 239 fuzz entry points across 123 OSS-Fuzz libraries, covering memory-safety errors and failing assertions under a modeled Linux OS response and memory layout. Anthropic calls this, to its knowledge, the first large-scale effort to use AI to apply formal verification to production software at the assembly code level. Anthropic says it is publishing a subset of the Rocq proofs and trusted components, and that Theorem is offering to run the same pipeline for other open-source projects and is publishing a Claude skill to help developers get started.
Anthropic also states clear boundaries. The proofs assume a single-threaded static binary with a fixed memory budget on a RISC-V CPU using Linux. Trusted artifacts include the formal models of RISC-V and Linux processes, the exact theorem statements, and the Rocq proof checker. Anthropic writes that this does not mean the libraries contain no bugs: remaining issues may sit in those trusted components or outside the proof scope. In short, LLM-scaled formal verification can prove a stated no-crash spec under those assumptions; it cannot, by itself, certify zero bugs in every deployment. For prior DigIntoAI notes on Anthropic security and agent tooling, see Anthropic Claude attack surface reduction, Anthropic Agent Skills to AAIF, and METR OpenAI agent monitoring red team.
We have not reproduced the Rocq proofs, re-run the OSS-Fuzz entry points, or verified Theorem’s Claude skill.
