XNews

Google DeepMind Agents Prove Signal libsignal Security Properties in Lean

Google DeepMind Agents Prove Signal libsignal Security Properties in Lean

Google DeepMind used Gemini-powered AI agents to prove key security properties of Signal's unmodified libsignal code down to the compiled machine code, writing about 6.5 million lines of Lean proofs. A team of five completed the proof work in four weeks, compared with twenty person-years to verify the seL4 microkernel. When two old Signal bugs were reintroduced, the system proved the buggy versions insecure.

Event version 1.

Key insights

Agents wrote millions of Lean proof lines

Agents powered by Gemini 4 Carbon wrote about 6.5 million lines of Lean proofs on the unmodified libsignal code. The proofs cover key security properties down to the compiled machine code.

Five-person team finished in four weeks

A team of five completed the proof work in four weeks. Verifying the seL4 microkernel took twenty person-years.

Humans and agents split specification and proof work

Humans and agents write the security specs, prover agents write the proofs, and Lean checks every proof before it is accepted.

Reintroduced bugs were proved insecure

When two old Signal bugs were reintroduced, the system proved the buggy versions were insecure.

Source posts