A four-week project generated approximately 6.5 million lines of checked proofs, with human experts defining the security claims and AI agents producing the proof code
Google DeepMind researchers used AlphaProof Code to prove specified security properties of a compiled libsignal implementation.
Google DeepMind researchers have used AI agents to mathematically prove key security properties of libsignal, the reference implementation of the Signal messaging protocol, connecting cryptographic guarantees to compiled machine code.
The work uses AlphaProof Code, a framework that combines human-directed specification writing with teams of agents that generate proofs checked by Lean, a programming language and proof assistant.
The researchers describe the result as the first verification of protocol-level security properties for a widely deployed, unmodified cryptographic protocol implementation at this depth. Their paper, published on November 10, examines the code within a defined execution environment and sets out the assumptions and limitations behind its guarantees.
Five researchers directed the proof work over four weeks in October 2026. The paper says none of the authors wrote the proof code themselves.
Connecting security guarantees to the compiled implementation
The team worked with libsignal’s production Rust code, compiling it within a command-line program for AArch64 Linux, an Arm-based computing environment.
The verification covers the protocol’s PQXDH handshake and Triple Ratchet components. The security properties examined include authentication, forward secrecy and post-compromise security, addressing how messages remain protected under specified scenarios involving exposed keys or compromised devices.
Much previous protocol analysis has examined mathematical designs, specially written implementations or narrower properties of production software. AlphaProof Code’s research connects security statements to the compiled implementation, where differences between a design and the code that executes it can introduce vulnerabilities.
The researchers used two established approaches. A symbolic model treats cryptographic operations abstractly, while a computational model bounds an attacker’s chances of success under assumptions about the underlying cryptography.
To test whether the specifications could identify insecure behavior, the team reintroduced two previously fixed libsignal bugs into separate experimental builds.
In both cases, the agents produced Lean-checked proofs that the altered implementation violated the symbolic security theorem. These were tests using historical bugs, rather than announcements of newly discovered vulnerabilities in the current software.
Ehren Kret, Chief Technology Officer of Signal, says: “As AI-driven security threats grow, having the core Signal protocol rigorously verified with automated theorem proving gives us even greater confidence that our code delivers on its privacy guarantees.”
Human researchers decide what must be proved
The paper places substantial responsibility on the people writing and reviewing the specifications. Those statements determine what security means for the program, which attacks the proof excludes and which assumptions it relies on.
Researchers developed the specifications through repeated exchanges with agents powered by Gemini 4 Carbon and Claude Opus 5.5. Agents reviewed drafts against existing research and the implementation, while human experts directed revisions and approved the statements.
Gemini 4 Carbon then powered the teams producing the proofs.
A separate checking process used Lean to validate each proof and confirm that it established the approved specification. This was designed to prevent agents from claiming success after weakening a requirement or changing a definition.
The paper describes why that control is necessary: an apparently valid proof can establish an inadequate claim if the specification has been made too restrictive or too weak.
Swarat Chaudhuri, a Research Scientist at Google DeepMind in London and Professor at The University of Texas at Austin, linked the project to more than two decades of work on formal software verification.
Writing on LinkedIn, Chaudhuri said: “Proofs are now largely automated, but human experts will need to decide what is proved with AI assistance.”
The authors report that most of their effort went into specifications. They also designed the agents’ working procedures, monitored progress and reviewed proposed changes to trusted assumptions.
Alongside the approximately 6.5 million lines of proof code, the work rests on roughly 26,000 lines of trusted specifications and machine-code definitions. Confidence in the result therefore depends on those foundations and Lean’s checking system, as well as the generated proofs.
A defined verification target and further research ahead
The verified target was the researchers’ compiled libsignal command-line program. The complete Signal applications, including their surrounding application behavior, were outside that verification scope.
Some low-level cryptographic and supporting routines were represented through contracts: formal descriptions of how those routines are assumed to behave. The proofs do not independently establish the correctness of those implementations.
The paper also specifies restrictions on the execution environment and attacker models. Its computational results do not provide guarantees against quantum attackers, while the symbolic analysis addresses some quantum-related handshake properties under its own assumptions.
Google DeepMind identifies medical device firmware and power-grid software as potential future applications of the approach, while acknowledging that scaling it requires further research. It also points to possible use alongside CodeMender, its AI agent for finding and fixing software security problems.
The paper’s nine authors are affiliated with Google DeepMind and Google. Signal researchers provided feedback on the methodology.
The team has released its Lean specifications and proofs for community use and plans to work with open-source contributors to improve them.
ETIH Innovation Awards 2027 open for entries
The ETIH Innovation Awards 2027 are now open for entries, with 26 categories recognising companies, institutions, projects and individuals making a measurable impact across education technology. The global awards span AI in education, digital learning, higher education, skills and workforce, inclusion, education operations, evidence and impact, with entries managed through ETIH’s new online awards platform.
Entries are open worldwide, with the early-bird deadline on Friday 18 December 2026 and the final entry deadline on Friday 26 February 2027 at 23:59 UK time.
Explore the 2027 categories and judging criteria, read the How to Enter guide or start an entry through the ETIH Innovation Awards 2027 platform.