Signal scrutinized by AI with cryptographic proofs

Google DeepMind claims to have formally verified part of Signal’s security using AI agents. The work focuses on libsignal code, not the entire app, aiming to show that human-checked proofs can be produced.

Google DeepMind claims to have used artificial intelligence agents to formally verify security properties of Signal, this time tackling the compiled code of libsignal rather than a simple theoretical model. The team says it validated the PQXDH handshake and the Triple Ratchet, two core blocks of the messaging protocol, with machine-read and human-checked proofs. The challenge is clear: to know if AI can help rigorously demonstrate that encryption code does what it promises. This is not the kind of issue solved with a simple “all is well.”

Une preuve sur le code réel, pas seulement sur le papier

The most interesting point is not the announcement itself, but the chosen ground. The researchers say they went beyond an abstract model of the Signal protocol to verify a compiled implementation of libsignal, specifically libsignal crates v0.105.4 and SPQR v1.6.0, assembled in a statically linked Linux AArch64 executable. According to their paper, this is the first protocol-level security proof for unmodified code of a widely deployed cryptographic implementation [to be verified].

The work relies on AlphaProof Code, a multi-agent framework that transforms security goals into mathematical specifications, then into Lean proofs verified by the language kernel. In this setup, humans choose the properties to prove, resolve ambiguities, and validate trade-offs, while agents generate the proof code. In other words, AI does not decide alone the meaning of the demonstration; it executes part of the mechanics. It’s less glamorous than a robot that “understands” cryptography, but much more useful.

Ce que la vérification couvre, et ce qu’elle laisse de côté

The verification covers one-to-one session code, with the PQXDH handshake and the Triple Ratchet mechanism. The researchers say they formalized four guarantees: forward secrecy, which protects past messages if the current state is compromised; authentication; post-compromise security, i.e., the ability to regain security after a compromise; and handshake secrecy. The paper adds two analysis parts, one symbolic and one computational, to frame cryptographic assumptions.

However, the result does not certify Signal as a whole. The scope excludes key generation, group messaging, multi-device side effects, and side channels. Some cryptographic blocks also rely on trust contracts, as well as standard Lean axioms and AArch64 model assumptions. The authors therefore specify that the proof provides strong assurance on a specific scenario, not a general blank check on the entire application. At this stage, the word “formal” does not erase scope limitations.

Une méthode lourde, mais déjà moins héroïque qu’avant

The workload gives an idea of the ambition. Five researchers worked for four weeks, with agents running continuously on their workstations. The systems produced about 6.5 million lines of Lean proofs verified by machine, backed by some 26,000 lines of specifications and AArch64 semantics considered trustworthy. The authors also say they wrote no proof lines themselves; they orchestrated production, review, and integration.

The potential gain is here: formal verification has long resembled an endurance sport, reserved for very costly projects. The researchers recall that the seL4 microkernel proof took 20 person-years, while CompCert required six years of development. Their own work does not cancel this history, but suggests that with agents, the cost could decrease. It remains to be seen if the same method holds on other codebases, as AI sometimes likes precision as much as a cat likes water.

Pourquoi Signal et pourquoi maintenant

The choice of Signal is no accident. The app has become a reference for encrypted messaging, and its cryptographic architecture serves as a basis for privacy guarantees closely watched by privacy advocates and security researchers alike. In this context, formally proving part of its behavior has industrial signaling value in the strong sense: it shows what agents can produce when the target is neither an academic puzzle nor a toy model.

Signal, for its part, mainly sees this as a trust reinforcement. Its CTO, Ehren Kret, believes the work “gives us even greater confidence in our code’s privacy guarantees.” The researchers present this step as a method test to go further: more verified routines, fewer trust assumptions, and eventually broader coverage of libsignal versions before release. For now, the idea is not to announce immediate deployment, but to show that a human-in-the-loop workflow can scale without losing control of assumptions.

To read