Skip to content
AI.info

The Pulse

DeepMind uses AI to verify key properties of Signal protocol code

Google DeepMind researchers used AI agents and human-reviewed specifications to verify security properties of compiled Signal protocol code. The proofs cover one-to-one sessions but have limits, including on claims about quantum attackers.

DeepMind uses AI to verify key properties of Signal protocol code

AI agents working with human researchers have formally verified key security properties of compiled code implementing the Signal messaging protocol. In a paper published November 10, Google DeepMind and Google researchers say the proofs are the first to connect protocol-level security guarantees to the unmodified code of a widely used cryptographic protocol implementation.

The researchers used a framework called AlphaProof Code to verify libsignal, the protocol’s reference implementation, generating about 6.5 million lines of machine-checked proof code in the Lean language. The team says the proofs address properties including forward secrecy, authentication, post-compromise security and handshake secrecy.

What the proofs cover

The verified program was an AArch64 executable built from unmodified libsignal v0.105.4 and SPQR code. The work concerns one-to-one sessions; group messaging is left for future work.

Variants of the Signal protocol are used in the Signal app, WhatsApp, Meta’s Messenger and Google Messages. The proofs cover the protocol’s PQXDH handshake and Triple Ratchet, components designed to protect messages even when some keys are compromised.

The researchers examined the compiled machine code rather than limiting their analysis to an abstract protocol design. They say the proofs establish security properties directly for the compiled binary against classical attackers, and handshake secrecy against a future quantum attacker in a symbolic model.

Humans set the specifications

AlphaProof Code combines a human-guided agent that writes specifications with autonomous agents that produce proofs. A specification states the properties to be proved and the assumptions about how code executes; researchers review and approve it before proof work begins.

A gatekeeper accepts a proof only if Lean’s checking system validates it and a separate comparison confirms that it proves the approved specification. For this project, human researchers chose the scope and reference models, guided the proof strategy and reviewed issues raised during the work.

The researchers say they did not write the proofs themselves. Google DeepMind says the agents that generated them were powered by Gemini Carbon. The paper reports that a team of five carried out the proof work over four weeks, running agents on their workstations.

Testing the boundaries

To test whether the specifications could catch real errors, the researchers say they reintroduced two previously fixed libsignal bugs. The resulting binaries violated the symbolic security theorems, according to the paper.

In one of the bugs, a prekey message that decrypted only under an archived session caused libsignal to store the new session’s state, including a copy of a secret. The authors report that the reintroduced bug made the resulting binary violate their symbolic security theorem.

The researchers used both symbolic and computational security models. The symbolic model treats cryptographic primitives as abstract components with standard security properties; the computational model instead bounds an attacker’s success in relation to the difficulty of breaking those primitives.

Limits of the result

The proofs rely on assumptions, including trust in Lean’s checking system, the security of cryptographic primitives and contracts for some routines called by the executable. The specification itself also requires review: the paper notes that a flaw in it could weaken the guarantee a theorem provides.

The computational proof gives no guarantee against a quantum attacker, although the symbolic proof covers part of that gap. Neither proof establishes the post-quantum healing that SPQR is designed to provide, the authors say. The computational model also excludes multi-device state and uses fixed message lengths.

The researchers note that specifications for widely deployed systems will need open review and refinement by the communities that maintain and rely on them. They are open-sourcing the Lean specifications and proofs for that process.

Sources

Companies in this story

Related stories

Explore

More articles