⏱️ Reading time: 9 min
An artificial intelligence system has just proven with mathematics, not intuition, that Signal’s encryption does exactly what it promises. Google DeepMind announced that its AlphaProof Code framework formally verified the security properties of the real code running behind the Signal protocol, the cryptographic foundation used by billions of conversations on Signal and Google Messages. The procedure, which historically consumed years of expert work, took just four weeks.
📑 En este artículo
- TL;DR
- What AlphaProof Code Is
- What Happened
- Context: Decades of Incomplete Verification
- Technical Details: How the Formal Verification Framework Works
- Impact and Analysis: What Changes for Software Security
- What’s Next
- Frequently Asked Questions
- What is AlphaProof Code and who developed it?
- Why did DeepMind’s tool choose the Signal protocol as its first case?
- How long did the Lean proof system take and how much proof did it generate?
- How was it confirmed that agentic code verification actually detects flaws?
- Does the formal verification framework make the Signal protocol infallible?
- Will DeepMind release the code for the Lean proof system?
- References
TL;DR
- Google DeepMind’s AlphaProof Code formally verified the real code of the Signal protocol in four weeks.
- Agents powered by Gemini Carbon generated more than six million lines of proof in the Lean language.
- Only 1% of the code needed direct human review; Lean automatically verified the rest.
- The system flagged as insecure two versions with real bugs that Signal had already fixed.
- Signal confirms that the verification increases confidence in the privacy guarantees of its protocol.
What AlphaProof Code Is
AlphaProof Code is a formal verification framework created by Google DeepMind that combines language-model-based agents with the Lean theorem prover to mathematically prove that a program meets its security specifications down to the level of the code that actually runs on devices, not just its paper design.
Its first public deployment was precisely the Signal protocol, long considered a benchmark challenge in security software. The result, according to DeepMind, is the first time the real implementation of a widely used cryptographic protocol has been formally verified, with or without artificial intelligence involved.
What Happened
Google DeepMind researchers guided artificial intelligence agents to prove, using the formal verification framework, the key security properties of the code underlying the Signal protocol, used by both the Signal app and Google Messages.
The work began with an assisted manual translation. Researchers used a combination of models to convert the security goals into formal mathematical specifications, and human experts reviewed each one before moving forward. From there, agents powered by the Gemini Carbon model took over generating the proofs, producing more than six million lines of code in Lean, the language that allows mathematically proving that a program is correct.
After generating the proof, the agents themselves audited it and flagged which parts needed a second human look: just 1% of the codebase. The rest was automatically validated by Lean, which checks every logical step with no room for error. According to DeepMind’s announcement, the entire process took four weeks, compared to the years expert teams usually need for a verification of this depth.
To confirm that the system can distinguish correct code from flawed code, the team deliberately reintroduced two real bugs, previously fixed, into an isolated version of Signal’s code. In both cases, AlphaProof Code generated a machine-verified proof that the buggy version was insecure.
“Using AI agents to translate real, live code into valid Lean theorems is a major breakthrough. As AI-driven threats grow, rigorously verifying the implementation of the Signal protocol gives us even more confidence in the privacy guarantees of our code,” said Ehren Kret, Signal’s Chief Technology Officer, according to DeepMind’s blog.
Context: Decades of Incomplete Verification
Cryptographic security has been analyzed almost exclusively on paper for decades. Researchers prove that a mathematical design is correct, but rarely manage to confirm that the actual implementation, the code that compiles and runs on a phone, follows that design to the letter. That’s usually where vulnerabilities hide: in the gap between the specification and the binary.
The Signal protocol has been the benchmark for end-to-end encrypted messaging for more than a decade. It’s used by Signal, Google Messages and, in different variants, other encrypted chat apps. That’s why it had a reputation as a benchmark challenge for formal verification: nobody had managed to mathematically prove that its code, not just its design, meets guarantees like keeping messages protected even after a device is temporarily compromised.
Technical Details: How the Formal Verification Framework Works
The system divides the verification work the way a human research team would. First, humans and models collaborate to set the specifications, that is, what must be true about the code. Then, a group of artificial intelligence agents, in this case powered by Gemini Carbon, focuses exclusively on producing the mathematical proof in Lean. Every step of that proof passes through the Lean verifier, which rejects any invalid logical step, so there’s no way for a proof to pass if it contains an error, because the language itself prevents it.
sequenceDiagram
participant H as Humans
participant M as Models
participant A as Gemini Carbon Agents
participant L as Lean
H->>M: define security goals
M->>H: proposes formal specification
H->>A: approves reviewed specification
A->>L: generates 6M+ line proof
L-->>A: verifies each logical step
A->>H: flags 1% for human reviewThis work stands out from previous formal verifications because of its depth: it went all the way down to the compiled code instead of stopping at the paper design. That choice multiplies the complexity. Real vulnerabilities almost always show up in the gap between the promised design and what the compiler ultimately produces. According to DeepMind’s announcement, the previous formal verification approach had necessarily been limited to smaller scopes, while this new method allows for significant expansion, as Ehren Kret, Signal’s Chief Technology Officer, explained in statements published by DeepMind.
The cost of this approach isn’t just computational: it’s about targeted expert review. Instead of a team of cryptographers reading all six million lines of proof line by line, Lean automatically validates the vast majority, and experts focus their time exclusively on the 1% that the agents themselves flagged as uncertain. It’s the same logic as a linter, but applied to mathematical proofs instead of code syntax.
Impact and Analysis: What Changes for Software Security
DeepMind frames this work as a proof of concept that extends beyond Signal. If an agentic verification framework can mathematically prove the security of a real cryptographic protocol, the same technique could extend to equally critical systems, such as medical device firmware or software that controls the power grid, where an implementation error has physical, not just digital, consequences.
The company also links this work to its automated defense agents, such as CodeMender, under the idea that formal verification and automatic vulnerability detection can be combined to keep pace with AI-driven cyber threats. DeepMind announced it will open-source the verification artifacts generated in this work, though it hasn’t yet specified under what license or when they will be fully available.
The announcement itself acknowledges an important limitation. Scaling this technique to other systems requires ongoing research, and it isn’t a button any engineering team can press today on their own codebase. The process required four weeks of coordinated work among Lean specialists, cryptographers, and the models themselves, not an unsupervised, end-to-end automatic run.
What’s Next
The next step DeepMind has stated is to open the artifacts from this verification to the community, so other cryptography teams can audit it or reuse the approach on their own protocols. Signal is already working with DeepMind to refine the methodology, as Kret confirmed, in the face of new threats emerging from AI agents capable of automating attacks at scale.
It remains to be seen whether other widely used protocols, such as TLS 1.3 or the post-quantum encryption schemes NIST standardized in FIPS 203, go through a similar process. In the meantime, the immediate point of comparison is that this work took four weeks where years were previously the norm, a shift in scale that, if repeated in other projects, would move formal verification from an academic luxury to a routine engineering practice.
Try it yourself: open the official Lean repository and see how a real proof is written before assuming formal verification is just academic theory.
Frequently Asked Questions
What is AlphaProof Code and who developed it?
It’s an agentic formal verification framework developed by Google DeepMind. It combines language models with the Lean theorem prover to mathematically prove that a program’s real code meets its security specifications.
Why did DeepMind’s tool choose the Signal protocol as its first case?
Because it had a reputation as a benchmark security challenge. It protects billions of conversations on Signal and Google Messages, and its real implementation had never been formally verified, only its paper design.
How long did the Lean proof system take and how much proof did it generate?
DeepMind reports that the verification took four weeks and that the agents produced more than six million lines of proof in Lean, of which only 1% of the code needed human review.
How was it confirmed that agentic code verification actually detects flaws?
The team reintroduced two real bugs, previously fixed in Signal, into an isolated version of the code. In both cases, the framework generated a Lean-verified proof that that version was insecure.
Does the formal verification framework make the Signal protocol infallible?
No. The proof covers the specific security properties that experts formalized as the specification, such as protecting messages before and after a temporary device compromise. An error in the original specification, or an issue outside that scope, wouldn’t be caught by this method.
Will DeepMind release the code for the Lean proof system?
Yes, the company announced it will open the verification artifacts for community use, though it hasn’t yet specified a license or full availability date.
References
- Google DeepMind: official announcement of AlphaProof Code and its formal verification of the Signal protocol.
- Signal: official site of the encrypted messaging app whose protocol was verified.
- Lean Prover: official documentation for the language and theorem prover used for the proofs.
- Wikipedia: historical and technical context on the Signal protocol.
📱 Like this content? Follow @programacion on Telegram for daily tech content in Spanish: quick summaries, fresh content every day.
Featured image: Foto de Jake Walker en Unsplash
How each article is made (sources, testing and editorial criteria): How we work.
Did it work for you? Got a different error? Say so below: questions get answered and help the next reader.
Leave a comment
0 Comments