Markets
BTC $56,195 -8.74% ETH $1,571 -9.04% BNB $517 -8.93% XRP $0.8799 -8.71% SOL $66.06 -8.42% LINK $8.14 -7.40% BTC $56,195 -8.74% ETH $1,571 -9.04% BNB $517 -8.93% XRP $0.8799 -8.71% SOL $66.06 -8.42% LINK $8.14 -7.40% TOTAL $1.97T BTC.D 57.2%

Blockchain Technology

AI Agents Wrote 6.5 Million Lines of Proof for Signal’s Encryption Build

Google researchers say a human-guided agent framework produced machine-checked proofs for one compiled build of the messaging library, with key generation left outside the scope.

Google researchers say their AI agents produced mathematical proofs tying specific security guarantees to a compiled build of the library that implements the Signal protocol, and they published the evidence so others can check it. The November 10 technical report describes about 6.5 million lines of machine-checked Lean proof code.

The target was narrow by design. Verification covered an AArch64 Linux executable built from libsignal-protocol v0.105.4 and SPQR v1.6.0 inside a small command-line harness. The theorems cover three operations: initialising a session, sending a message and receiving one. Key generation sits outside the verified scope, as do the phone applications and group messaging.

The work was directed by people. Five researchers spent four weeks selecting reference models, writing and reviewing the specifications that defined what the agents had to prove, and checking the output. The agents then wrote proofs against those specifications, and a comparator gate confirmed that each proof matched the stated theorem.

What was proved concerns the post-quantum handshake and the ratcheting schedule used for one-to-one messaging, including forward secrecy, authentication and post-compromise security. Those properties answer three practical questions: whether earlier messages stay secret after keys are stolen, whether a sender is who they claim to be, and whether a session can recover secrecy after a compromise. Two proof models were used. The symbolic analysis assumes idealised cryptographic operations and collision-free execution, while the computational analysis bounds the success of classical attackers under stated assumptions.

The limits are as important as the result. A checked proof establishes the stated guarantee for that build, that specification and that execution model. The specifications, the machine model, routine contracts and the cryptographic assumptions all remain trusted inputs rather than things the proofs establish. A change to the build or to the specification invalidates the guarantee without invalidating the method.

For anyone running cryptographic infrastructure, that framing is the substance of the story. Formal verification has been available for years to teams willing to hand-write specifications and proofs, and the cost has kept it out of reach for most production code. Cutting the proof-writing from months of specialist labour to four weeks of five people directing agents changes the economics, even if it leaves the trusted base untouched.

Whether the same approach holds for code that handles keys, rather than code that handles messages, is the next question. The report does not answer it.

Share this article

Viber uses viber://forward?text= when the app is installed.