GenProof, a new research lab, says it has found bugs in more than 100 open-source projects, including six libraries that Anthropic had described as "formally verified." The team also claims the top spot on the lean-eval math formalization benchmark and is releasing its winning harness. The lab is launching QuickProof, a tool designed to let developers who can write unit tests perform formal verification. The approach uses an LLM to generate proofs that properties like unzip(zip(data)) == data hold, with proofs machine-checked so LLM output is never trusted. The author attributes multiple bugs and CVEs to this method.

AI summary of the linked source — not editorial content.

Explore the topic

Loading related coverage…

Explore the full feed

Comments

Loading comments…