RESEARCH · RESEARCH · #3323
DeepMind's AlphaProof Code formally verifies parts of the Signal protocol
Google DeepMind used a new agentic AI framework called AlphaProof Code to formally verify several key cryptographic properties of the Signal protocol in a four-week proof-of-concept project. The analysis found no new vulnerabilities in the examined portions, confirmed that older known-vulnerable versions fail verification, and released the specification and proof, while researchers cautioned that scaling AI-driven formal verification remains an open challenge.
KEY POINTS
- Google DeepMind used a new agentic AI framework called AlphaProof Code to formally verify several key cryptographic properties of the Signal protocol in a four-week proof-of-concept project.
- The analysis found no new vulnerabilities in the examined portions, confirmed that older known-vulnerable versions fail verification, and released the specification and proof, while researchers cautioned that scaling AI-driven formal verification remains an open challenge.
- This demonstrates that AI can materially accelerate formal verification of widely used cryptographic protocols, validating Signal's guarantees while highlighting open questions about scalability and practical deployment.
DECISION BRIEF
CONFIDENCE
WHAT CHANGED
Google DeepMind used a new agentic framework called AlphaProof Code to produce formal mathematical proofs for several key cryptographic properties of parts of the Signal protocol and published the resulting specification and proof artifacts.
WHY NOW
This project demonstrates that agentic AI can materially accelerate producing formal proofs for widely deployed cryptographic code, reducing human review to a small fraction of the code in this proof-of-concept—raising immediate questions about whether the approach can be scaled and adopted more broadly.
WHO IS AFFECTED
Directly affected: DeepMind researchers and the Signal protocol codebase portions analyzed; indirectly affected: users and vendors relying on Signal protocol implementations (e.g., Signal app, Google Messages, WhatsApp) who depend on the verified properties.
CONFIRMED
- DeepMind built an agentic framework called AlphaProof Code and used it to formally verify several key cryptographic properties of parts of the Signal protocol (DeepMind blog). - The proof work was done as a proof-of-concept in ~four weeks (WIRED). - Agents (including models referred to as Gemini Carbon) generated over six million lines of Lean and produced the proofs; humans reviewed specifications and checked agent output (DeepMind blog). - Agents flagged about 1% of the codebase for additional human review; the remainder was checked by Lean (DeepMind blog). - The analysis did not find new vulnerabilities in the portions examined (WIRED). - Applying AlphaProof Code to older Signal versions with known vulnerabilities correctly showed those versions failed verification (WIRED). - DeepMind published/shared the specification and proof artifacts from the project (WIRED; DeepMind blog language indicates public sharing).
UNCERTAIN
Missing or unresolved items in the supplied excerpts: - Exact boundary/scope of the 'parts' of Signal that were verified (which modules, how much of the codebase) is not specified in these excerpts. - Degree to which the verification covers the full compiled runtime on user devices vs selected implementation components: the blog claims verification down to the compiled code, while WIRED emphasizes the analysis was limited and did not verify the entire mobile app — this is a scope tension. - Reproducibility and independent third‑party validation beyond DeepMind's publication are not reported here. - Practical scalability and resource costs for applying AlphaProof Code to much larger codebases remain open (sources state this as an open challenge). - Any limitations in the formal properties proven (which exact security properties were proven and which were not) are not fully enumerated in the excerpts. (These points are explicitly missing or presented as open questions in the provided sources.)
WHAT TO WATCH
Observable signals to monitor next: - Release locations and contents of the Lean files and full specifications/proofs (public repo, DeepMind paper or dataset). - Independent third-party audits or reproductions using the published proofs/specification. - Follow-up publications or technical writeups from DeepMind detailing scope, methodology, and resource costs. - Responses or technical assessments from the Signal Foundation or other implementers about coverage and limitations. - New projects applying AlphaProof Code to larger portions of Signal or to other widely used protocols and any reported human-review percentages or lines-of-proof metrics.
WHY IT MATTERS
This demonstrates that AI can materially accelerate formal verification of widely used cryptographic protocols, validating Signal's guarantees while highlighting open questions about scalability and practical deployment.
EVIDENCE MAP
4Editorial claims linked to specific sources, with support, contradiction and context shown separately.
DeepMind used a new agentic framework called AlphaProof Code to formally verify several key cryptographic properties of the Signal protocol.
SUPPORTEDChecked 2026-11-10 · 2 supporting
WIRED reports DeepMind used agentic AI to mathematically prove a number of the Signal protocol's key guarantees using a framework called AlphaProof Code.
SUPPORTS Google DeepMindMEDIA · 2026-11-10DeepMind's blog excerpt says researchers 'guided AI agents to mathematically prove key security properties' of Signal using AlphaProof Code.
WIRED describes the work as a proof-of-concept that was conducted over four weeks.
SUPPORTEDChecked 2026-11-10 · 1 supporting
WIRED explicitly says the work was conducted as a proof‑of‑concept in just four weeks.
CONTEXT Google DeepMindMEDIA · 2026-11-10DeepMind's excerpt frames the effort as a milestone and proof that agents performed deep formal verification, but does not state the four‑week duration in the supplied excerpt.
The analysis did not find any new vulnerabilities in the portions of the Signal protocol that were examined.
SUPPORTEDChecked 2026-11-10 · 2 supporting
WIRED states DeepMind 'did not find any new vulnerabilities in the portions of the Signal protocol it looked at.'
SUPPORTS Google DeepMindMEDIA · 2026-11-10DeepMind's excerpt likewise reports the analysis found no new vulnerabilities in the examined portions (as presented in the supplied excerpt).
When run on older Signal versions that contain known, fixed vulnerabilities, AlphaProof Code correctly showed those versions could not be formally verified.
SUPPORTEDChecked 2026-11-10 · 2 supporting
WIRED reports that DeepMind used AlphaProof Code on older Signal versions with known vulnerabilities and showed the protocol could not be formally verified.
SUPPORTS Google DeepMindMEDIA · 2026-11-10DeepMind's excerpt includes the claim that older known‑vulnerable versions fail verification (as presented in the supplied excerpt).
SOURCES & TIMELINE
3How Google DeepMind researchers built a new agentic framework that formally verified the code of Signal’s cryptographic protocol — long considered to be a grand security challenge Today, we’re sharing a milestone in software security long considered to be a grand security challenge: Google DeepMind researchers guided AI agents to mathematically prove key security properties of the actual code underlying the Signal p…
Google DeepMind researchers used agentic AI to mathematically prove, or “formally verify,” a number of the Signal protocol’s key cryptographic guarantees—a landmark development for the widely used cryptography and AI itself. Conducted as a sort of proof-of-concept project in just four weeks, the work is significant in two ways. It serves as further independent validation of the cryptography that protects massive qua…
Google Deepmind has unveiled AlphaProof Code, an AI system that mathematically proves software meets its security specs without errors. Researchers demonstrated the technology on Signal's encryption protocol, completing in four weeks work that typically takes years. Earlier verification efforts stopped at the theoretical design stage, according to a blog post by DeepMind researchers Pushmeet Kohli and Swarat Chaudhur…