Formal Verification of the Signal Protocol Using AI
The researchers used formal verification to check whether parts of Signal’s protocol meet mathematical descriptions of its security goals. This matters because the protocol protects communications on platforms including Signal, Google Messages, and WhatsApp. But the article does not name the individual guarantees they verified, so it cannot support a more specific list. The work began by translating Signal’s security goals into formal mathematical specifications. Human experts reviewed each specification, then AI agents generated proofs in Lean. The result was a mathematical check of the stated goals for the portions examined, not a report identifying every security property of the entire protocol. The researchers found no new vulnerabilities in the parts they analyzed. They also tested older Signal versions with known, long-fixed vulnerabilities, and the system flagged that those versions could not be formally verified. The team is releasing its specification and proof to support other verification work. The article describes the project as a four-week proof of concept, not a complete verification of every part of Signal.
Which key security guarantees of the Signal protocol did the researchers formally verify?
The researchers used formal verification to check whether parts of Signal’s protocol meet mathematical descriptions of its security goals. This matters because the protocol protects communications on platforms including Signal, Google Messages, and WhatsApp. But the article does not name the individual guarantees they verified, so it cannot support a more specific list.
The work began by translating Signal’s security goals into formal mathematical specifications. Human experts reviewed each specification, then AI agents generated proofs in Lean. The result was a mathematical check of the stated goals for the portions examined, not a report identifying every security property of the entire protocol.
The researchers found no new vulnerabilities in the parts they analyzed. They also tested older Signal versions with known, long-fixed vulnerabilities, and the system flagged that those versions could not be formally verified. The team is releasing its specification and proof to support other verification work. The article describes the project as a four-week proof of concept, not a complete verification of every part of Signal.
What is the Signal protocol, and how is it used by apps such as Signal, Google Messages, and WhatsApp?
The Signal protocol is a cryptographic protocol: a set of rules and methods that apps can use to protect digital messages. In general, secure messaging protocols help make messages readable to their intended recipients rather than outsiders. The article describes Signal’s protocol as protecting communications on Signal, Google Messages, and WhatsApp.
For example, when people exchange messages through an app using the protocol, cryptographic methods can help protect those messages as they are sent. The article’s researchers focused on whether key security goals for the protocol could be expressed and proved mathematically. It does not explain the protocol’s specific technical steps or say which guarantees were included in their analysis.
The article calls the protocol widely used and says it protects massive quantities of communications. It reports that the researchers found no new vulnerabilities in the portions they examined. Because the article gives only a brief description of the apps’ use of the protocol, details about how each app implements it are not provided. In general, implementations can differ, so the article’s findings should be understood as covering the analyzed portions, not every app feature.
How large was the AI-generated proof effort, and how much of it did human experts need to review closely?
The AI-generated proof effort produced over six million lines of Lean, a computer language used to mathematically prove the accuracy of code. That scale is striking, but the article’s main point is not simply the amount of code: the process let human reviewers concentrate on a much smaller portion.
After human experts reviewed the formal specifications, agents powered by Gemini Carbon generated the proofs. The agents then audited the proof and highlighted areas that needed a second look. Those flagged areas amounted to just one percent of the codebase, according to the article. Lean automatically checked the rest, reducing the amount of code that experts needed to review closely.
The work was a proof-of-concept project conducted in just four weeks. The article presents the smaller review burden as a major time saving, while still giving human experts a role where judgment mattered. It does not provide a detailed breakdown of the six million lines or say how long a comparable entirely human effort would have taken. The researchers are releasing their specification and proof to aid further formal verification work.
What happened when the AI analyzed older Signal versions with known vulnerabilities?
The researchers tested AlphaProof Code not only on the Signal portions they were analyzing, but also on older versions of Signal known to contain vulnerabilities that had since been fixed. This provided a check of whether the AI-assisted approach could detect a failure rather than simply produce a proof in every case.
In those older versions, AlphaProof Code correctly identified the known vulnerabilities and showed that the protocol could not be formally verified. The article does not describe the vulnerabilities themselves or explain the exact proof failures. It reports the outcome as a successful identification of known problems, not as discovery of new ones.
For the portions of the current protocol examined, the researchers found no new vulnerabilities. They are releasing the specification and proof they developed, with the aim of helping other formal verification work. The article frames this as a four-week proof of concept and a demonstration of DeepMind’s verification capability. It does not claim that the method has now checked every version or every component of Signal.
How did AI agents, human experts, and Lean work together to check the protocol?
The process combined AI, human review, and Lean, a computer language that mathematically proves the accuracy of code. Its purpose was to check whether Signal’s stated security goals held for the protocol portions under analysis. The article presents this as a way to save review time without removing human judgment.
First, researchers worked with a mix of models to translate Signal’s security goals into formal mathematical specifications. Human experts reviewed every specification. Then agents powered by Gemini Carbon generated over six million lines of Lean proof code. The agents audited the proof and highlighted areas for a human second look; those areas made up just one percent of the codebase.
Lean automatically checked the remaining code, allowing experts to focus their review where human judgment mattered most. The researchers found no new vulnerabilities in the portions they examined, and the article calls the four-week effort a proof of concept. The team is releasing its specification and proof to aid other verification work. The article does not claim this process replaces expert review or verifies all of Signal.
What is formal verification, and how can a mathematical proof show that software meets a security specification?
Formal verification means using mathematics and specialized tools to check whether a system satisfies precisely stated requirements. It matters for security because vague claims such as “this software is safe” are hard to test directly. A formal specification instead describes a goal in a form that a proof checker can examine. The article says researchers used this method on key cryptographic guarantees of Signal.
For example, the team first translated Signal’s security goals into formal mathematical specifications, with human experts reviewing each one. AI agents then generated proof code in Lean. Lean automatically checks whether the proof follows from its formal rules. If a specification cannot be proved for a version, that signals the stated guarantee has not been established for that version.
The researchers report that older Signal versions with known vulnerabilities could not be formally verified, while their analysis found no new vulnerabilities in the portions examined. A proof applies to the specification and system parts actually modeled; it does not automatically show that all software or every real-world use is secure. The article describes a four-week proof of concept and says the specification and proof will be released.
How does cryptography help keep messages private and protected from tampering as they travel between people?
Cryptography uses mathematical methods to protect information. In messaging, encryption can make a message unreadable to people who do not have the right way to decrypt it. Other cryptographic methods can help detect whether information has been altered. These are general explanations; the article itself does not describe the specific mechanisms used by Signal’s protocol.
A simple example is a message transformed into an unreadable form before it travels between people, then restored for an intended recipient. A protocol sets rules for how such protections are used. The researchers in the article focused on expressing Signal’s security goals as mathematical specifications and proving them with Lean, rather than describing a particular message exchange.
The article says Signal’s cryptography protects massive quantities of communications on platforms such as Signal, Google Messages, and WhatsApp. The researchers found no new vulnerabilities in the portions they examined, and they are releasing their specification and proof to support further work. Cryptography can help protect privacy and integrity, but the article does not claim it prevents every possible security problem or explain the protections of each app in detail.
This brief was written by AI from the original reporting and checked by other models. Names, figures and quotes come from the source; read it for full context.
Read more in the JupiteX app
Pulse is free. New stories every 4 hours, each one broken into the questions that explain it.
Or read more news on the web