Skip to content
Quantum.world

The Signal

Google DeepMind says it verified Signal code at machine-code level

Google DeepMind researchers used AI agents to produce proofs about key security properties of Signal’s compiled cryptographic code. The work covers a post-quantum handshake, but does not prove the ratchet’s post-quantum recovery properties.

Google DeepMind says it verified Signal code at machine-code level

Google DeepMind researchers say AI agents have produced machine-checked proofs of key security properties of the compiled code behind the Signal protocol. Their work reaches the machine-code level rather than stopping at a description of the protocol’s design.

The proofs cover the post-quantum PQXDH handshake and the Triple Ratchet, the researchers report. They address properties including handshake secrecy, authentication, forward secrecy and security after a compromise, although the guarantees differ between the models used and have important limits against quantum attackers.

What was verified

The team verified an AArch64 executable built from the unmodified libsignal v0.105.4 and SPQR crates. The executable contains the reference implementation’s protocol code; the work concerns the compiled program rather than a modified version written for verification.

The handshake establishes a shared secret for a conversation, while the ratchet derives fresh keys for messages. The paper describes forward secrecy as protecting earlier encrypted messages when a current state is stolen, and post-compromise security as restoring protection after a compromise under specified conditions.

The verification treats cryptographic primitives, output functions and some library routines through trusted contracts. It also relies on the proof checker and a model of the machine code’s behavior. The work covers one-to-one sessions, not group messaging.

The quantum boundary

The researchers used symbolic and computational models to state and prove different security guarantees. In the symbolic model, cryptographic operations are idealized, and an attacker can later break the key agreement used to represent a future quantum attack. Under that model, the authors report a proof of handshake secrecy even against such an attacker.

The computational proof instead gives a bound on an attacker’s advantage under assumptions about the underlying cryptographic primitives. Its bounds do not give a guarantee against a quantum adversary, the paper states.

Neither set of ratchet results proves the post-quantum recovery that the sparse post-quantum ratchet is designed to provide. Other limits include fixed message lengths in the models and restrictions on the sessions and adversary covered by the computational proof.

Agents and human review

Researchers worked with models to turn security goals into formal specifications, with human experts reviewing each specification. Agents powered by Gemini 4 Carbon then generated the proofs, according to DeepMind.

The agents generated about 6.5 million lines of machine-checked Lean proof code, the researchers report. The proof checker accepts an agent’s work only when it checks against the approved specification; the agents also flagged material for further human review.

DeepMind says the verification process took four weeks. To test the specifications, the researchers reintroduced two previously fixed bugs into an isolated version of the code; in both cases, the system produced a machine-checked proof that the buggy version was insecure.

The specifications and proofs are available for community use. The researchers also identify further work on the trusted parts of the verification and on extending its scope to other parts of the messaging system.

Sources

  1. Google DeepMind
  2. Google DeepMind

Related stories

Explore

More articles