Skip to content
trendoviz

AlphaProof Code verifies Signal’s core cryptography

Google DeepMind used AlphaProof Code, a multi-agent formal verification system, to prove selected security properties of Signal’s compiled cryptographic implementation. Human experts defined and reviewed the specifications, AI agents generated Lean proofs, and the project team reported that Lean checked every proof.

55 out of 100Emerging7D: Insufficient history ProvisionalPublished

Overview

AlphaProof Code is a multi-LLM framework for formal verification. Human experts translate security goals into formal mathematical specifications and review each one, while AI agents generate and audit the proofs in Lean, a language for machine-checked mathematical proofs. The agents identified one percent of the codebase for a second human review, with Lean checking the remainder. Google DeepMind says every proof generated through the system was checked by Lean and that the formal verification artifacts are being open-sourced. [1]

Why it matters

Formal verification can provide stronger assurance than testing for selected security properties because a machine checks the proof rather than a tester sampling program behavior. AlphaProof Code points toward a division of labor in which people set the assurance requirements and AI agents handle repetitive theorem generation. [1]

The results reported so far are tied to Signal’s core cryptographic implementation and its specified properties. The evidence does not show that the system can verify arbitrary complex software, and Google DeepMind explicitly notes that scaling the technique requires further research. [1]

Opportunities

Google DeepMind presents the approach as a foundation for verifying other critical systems, including medical-device firmware and power-grid software. Those are proposed applications rather than demonstrated results. [1]

Developers may also be able to use the planned open-source verification artifacts as a reference for proof-oriented workflows. The evidence does not specify whether the release will include the specifications, Lean proofs, agent implementation, and bug-reintroduction setup, so their practical reuse is uncertain. [1]

Risks

Reliability and reproducibility remain open questions. The bug-reintroduction experiment was run by the project team, and the evidence provides no reproducible details about the two injected bugs or the corresponding proof artifacts. No independent third-party evaluation is identified. [1]

Coverage is also unclear. The exact scope and size of the verified codebase, the precise contents of the six million generated lines of Lean, and the proportion of the codebase covered by a complete proof are not stated. How often agents needed correction is also unknown. [1]

Organizations adopting the workflow would still need qualified experts to formulate specifications and review flagged areas, creating labor and skill dependencies. Google DeepMind’s comparison between four weeks and years of expert work has no comparative study, baseline, or named prior project in the evidence. Lock-in, legal, and cost effects are unknown. [1]

Context

Formal verification means proving properties of a program mathematically rather than relying only on testing and manual review. Google DeepMind says most prior protocol verification stopped at the design level, while this work examined the compiled implementation of Signal’s core cryptographic code. One specified property concerned keeping messages protected before and after a device is temporarily compromised. The work was conducted as a four-week proof of concept. [1][2]

The framework therefore does not remove human involvement. Experts select which properties to prove and review every formal specification; agents produce the proofs and flag areas for another human look. This is a workflow for proving chosen properties of a defined implementation, not evidence of general-purpose software verification. [1]

Known, claimed, inferred

Known: The work was a four-week proof of concept, according to the media account [2]. Google DeepMind describes AlphaProof Code as a multi-LLM framework in which human experts select security properties and AI agents generate proofs [1]. Human experts reviewed the formal specifications; agents reportedly generated more than six million lines of Lean; agents highlighted one percent of the codebase for a second human review, while Lean checked the rest [1]. The verification reportedly concerned compiled Signal core cryptographic code and included a property involving message protection before and after temporary device compromise [1]. Google DeepMind says all proofs were checked by Lean and that verification artifacts are being open-sourced [1].

Vendor claims: Google DeepMind says this was the first use of AI to formally verify a widely deployed cryptographic protocol at this depth, and the first formal verification of a widely deployed protocol implementation of this kind by any method [1]. It also says comparable work historically takes expert teams years, frames the result as a foundation for verifying medical-device firmware and power-grid software, and presents the work as assurance for the cryptography used across Signal, Google Messages, and WhatsApp [1][2].

Reported experiment: The project team reintroduced two previously fixed bugs into an isolated version of Signal’s code, and the system produced a machine-checked proof that the buggy version was insecure in both cases [1]. This is an account from the project team rather than a separately published third-party test report.

Inference: The workflow suggests that near-term value may come from reducing repetitive proof generation while retaining human control over specification and review, rather than from fully autonomous verification of entire codebases [1].

Limitations & uncertainties

There is no independent third-party evaluation, and no published comparative study supporting the four-weeks-versus-years claim [1][2]. The specific verified Signal properties are not fully listed beyond the temporary-compromise example [1]. The exact codebase scope and size, the contents of the generated Lean, and the proportion covered by a complete proof are unknown [1]. The injected bugs, their machine-checked artifacts, and an independent review of the results are not provided [1]. The evidence does not support verification of arbitrary complex software and explicitly notes that scaling requires further research [1]. It is unknown whether the planned open-source release includes specifications, Lean proofs, the agent implementation, and the bug-reintroduction setup [1]. The relationship between “Gemini Carbon” and the separately mentioned “Gemini 4 Carbon” is unexplained [1].

Evidence & sources

The key statements of this analysis, labeled by type, with the sources they rely on. Trendoviz does not run product tests.

  • Primary-source or vendor statementGoogle DeepMind used AlphaProof Code to guide AI agents toward mathematical proofs of key security properties in the compiled implementation of Signal’s core cryptographic code.[1]
  • Primary-source or vendor statementThe framework assigns specification work to human experts and proof generation to AI agents, with human experts reviewing every formal specification.[1]
  • Primary-source or vendor statementAgents reportedly generated more than six million lines of Lean and highlighted one percent of the codebase for a second human review.[1]
  • Primary-source or vendor statementGoogle DeepMind says every proof was checked by Lean and that the formal verification artifacts are being open-sourced.[1]
  • Primary-source or vendor statementThe project team reported that reintroducing two previously fixed bugs produced machine-checked proofs that the buggy version was insecure in both cases.[1]
  • Verified factA media account describes the work as a four-week proof of concept covering selected Signal cryptographic guarantees.[2]
  • Primary-source or vendor statementGoogle DeepMind says the technique historically would take expert teams years and could extend to medical-device firmware and power-grid software.[1]
  • HypothesisThe reported results apply to Signal’s core cryptographic implementation and do not establish that AlphaProof Code can verify arbitrary complex software.[1]

Sources

  1. 1
  2. 2
    Formal Verification of the Signal Protocol Using AI
    schneier.comretrieved Nov 18, 2026

Updates & corrections

  1. First published · published automatically after passing every check

Found an error? Report it.