SI Agent News
DevelopingGeminiAgentsSecurityResearch

DeepMind uses AI to mathematically verify Signal's encryption code

The new AlphaProof Code framework uses AI agents to prove that Signal's security guarantees hold at the machine-code level.

6.5 millionlines of machine-checked Lean proof code

The short version

Why it matters

This marks a shift from checking if code 'looks' correct to mathematically proving it cannot fail. It allows for the verification of complex, production-grade software that was previously too labor-intensive to audit manually.

Bridging the implementation gap

Google DeepMind recently unveiled AlphaProof Code, an agentic artificial intelligence framework designed for formal software verification. This process involves proving mathematically that a piece of software follows its intended specifications exactly. While this is the gold standard for cybersecurity, it has traditionally been too difficult and slow for most companies to use.

Previous landmarks in this field, such as the seL4 microkernel, required years of manual labor by human engineers to create proofs. AlphaProof Code changes this by using a multi-agent framework powered by Gemini Carbon models to synthesize these proofs at a much larger scale. The system relies on the Lean 4 proof assistant to ensure every step remains mathematically sound.

By combining machine learning with a deterministic kernel, the framework prevents the AI from making logical errors, which are often called hallucinations. This hybrid approach allows the system to propose complex proof tactics while the Lean assistant validates the final certified output. As reported by CASRAI, this allows for a rigorous approach to verifying software that was previously considered too complex for automated tools.

Verifying Signal's core security

The DeepMind team, led by Swarat Chaudhuri and Pushmeet Kohli, applied this technology to the Signal protocol. They specifically targeted the compiled machine code of the libsignal reference implementation and the spqr cryptographic crate. This is significant because it moves beyond abstract models to verify the actual instructions executed by the hardware.

Over a four-week period, human cryptographers worked with specification-writing agents to create roughly 26,000 lines of formal specifications. These specifications describe the rules for the ARM architecture and the requirements for the software. The AlphaProof Code agents then took over to generate the massive volume of proof code needed to verify the system.

The verification covers two essential parts of modern messaging: the post-quantum PQXDH key-agreement protocol and the Triple Ratchet mechanism. These components ensure that messages remain private even against future quantum computers. The proofs confirm that the system maintains message authentication, forward secrecy, and post-compromise security, which protects users if a device is temporarily hacked.

Detecting hidden vulnerabilities

To test if the system actually worked as a security tool, the researchers performed a controlled experiment. They intentionally put two old, historically patched vulnerabilities back into an isolated version of the Signal app. They wanted to see if the AI could identify these flaws as deviations from the formal rules.

AlphaProof Code successfully produced automated proofs showing that the flawed code failed to meet the required security specifications. This confirms that the framework can act as an active vulnerability detector rather than just a passive check. It can identify bugs that might be missed by traditional human code reviews or standard testing methods.

Despite the automation, the team maintained a human-in-the-loop boundary to ensure accuracy. The prover agents audited the generated proofs and flagged about 1% of the codebase where the automated synthesis stalled. This allowed human experts to focus their attention only on the most difficult and ambiguous edge cases in the code.

Impact on research and infrastructure

Signal Chief Technology Officer Ehren Kret noted that this agentic methodology allows for a much wider expansion of verification across production infrastructure. While previous projects only looked at small pieces of code, this framework can handle much larger systems. It provides a way to replace heuristic inspection, which relies on experience, with unconditional mathematical proof.

For academic and research institutions, this technology could soon be used to protect sensitive scientific software and data infrastructure. DeepMind has open-sourced the Lean formal specifications and verification artifacts to help the community. This allows computer science and cryptography faculties to use these results as reproducible benchmarks for automated reasoning.

As organizations look for ways to govern AI and machine learning research, these formal methods provide a way to prove the provenance of software pipelines. By preserving these proof libraries in archives like the Software Heritage archive, the team hopes to establish a new standard for scientific infrastructure. This move helps ensure that the software powering modern research is both reliable and mathematically verified.

What to watch

Try it

Check the Google DeepMind blog for the open-sourced Lean formal specifications to see the underlying math used to verify Signal.

Sources and links

Liked it? Say why

Add a line for other readers.

No account needed. We never show your email. Comments with links, ads or abuse are held and not shown. Your comment shows after a quick check.

Listen to the AI Wire headlines

Listen to the AI Wire

Listen to our daily AI shows every morning: on your commute, in the shower, wherever you are

Listen to the daily shows

First reported by CASRAI · Read the original. Written by the SI Agent News Desk from the reporting above.Written with AI from the source above. Every number is checked against the source. Spot a mistake? Corrections.

More AI news