Formal verification of x86 code: failed proofs led straight to real glibc bugs
moyix · x · 2026-11-20
Second post in the v12sec x86 formal verification project: formal verification doesn't find bugs—it proves none exist. But what happens when you try to prove code that has a bug?
While verifying four memchr variants in glibc, failing proofs pointed directly at a real bug: the code returns 0xFFFFFFFFFFFFFFFF instead of NULL at address 0. The accompanying chart traces Claude's 100-minute workflow across 8 proof runs—reading failed results, replaying goals, dumping open goals with a small tool, getting a counterexample region from omega, grepping the assembly—until adding a 32 ≤ s contract constraint made run 8 pass fully.
More from coding & agent
- Raycast launches MCP: let Claude, Codex and any agent drive your Mac launcher — dean_rie · 2026-11-20
- A 6-model board for decisions: hide opinions, let Claude chair and adopt the sharpest argument — aryanXmahajan · 2026-11-20
- Zed Delta agents can now run background terminals and get notified on completion — dean_rie · 2026-11-20
- Gemini app for Mac ships computer use, automating tasks across folders, apps and the web — GeminiApp · 2026-11-20
- Linear data: 53% of merged code now ships with AI review only, no second engineer — rseroter · 2026-11-20
- Claude builds its own tooling mid-task, cutting overnight jobs to one hour — DaniyarQQQ · 2026-11-20