November 20, 2026

Google DeepMind Uses AI Agents to Prove Signal’s Encryption Code Works as Intended

a close up of a cell phone with the google logo in the background

Google DeepMind has used AI agents to mathematically prove key security properties of the code behind the Signal protocol, the encryption technology that protects private conversations in Signal, WhatsApp and Google Messages.

The work, announced on 10 November, was carried out with AlphaProof Code, a new framework in which human experts and AI agents work together on formal verification. Formal verification means proving mathematically that software behaves exactly as specified, rather than testing it and hoping no flaws remain. It is considered the highest standard of security assurance, but it has traditionally been slow and labour-intensive.

The researchers checked libsignal, Signal’s open-source encryption library, at the level of the compiled machine code that actually runs on devices, rather than the source code. The properties they proved include forward secrecy, authentication and the secrecy of the initial key exchange, which also covers Signal’s post-quantum key agreement.

AI agents produced around 6.5 million lines of machine-checked proof in four weeks. Speaking to Axios, DeepMind researcher Swarat Chaudhuri said the same work would likely have taken human teams years. Human experts reviewed every specification, and the agents flagged about 1% of the code for an additional human check.

To test the system, the team reintroduced two bugs that Signal had previously fixed. In both cases, AlphaProof Code produced proofs showing that the altered code was not secure.

Signal’s chief technology officer, Ehren Kret, said that with AI-driven security threats growing, having the core protocol verified in this way gives the company greater confidence in its privacy guarantees.

The result covers the core of the protocol rather than the whole Signal app, and attacks such as timing and power analysis are out of scope. No new vulnerabilities were found. DeepMind describes it as the first time AI has been used to formally verify a widely deployed cryptographic protocol at this depth.

DeepMind expects the approach to be useful well beyond messaging. Pushmeet Kohli, DeepMind’s vice president of research, pointed to banks and other companies that rely on strong security assurances, which could use the method to say “I can prove these properties” about their own software.

Payment companies are a likely candidate. Many already use AI to detect fraud and monitor transactions, and checking the code that signs and sends payments would be a natural next step, especially in digital asset payments, where a mistake usually can’t be undone.

“One of the most sensitive parts of a crypto payment system is the code that signs transactions, generates wallet addresses and decides when funds can be withdrawn,” said Alex Ryzhkin, Chief Commercial Officer at B2BINPAY, a crypto payment provider that lets businesses accept, store and send payments in cryptocurrencies and stablecoins. “Today every change to that code goes through internal review and testing, and the platform is regularly audited by external security firms. With new blockchains and tokens being added all the time, the code changes often. AI could make it possible to re-run proofs automatically after every update. That wouldn’t replace human review, but it would add a layer of assurance that keeps up with how often the code changes.”

DeepMind has published its specifications and proofs openly, allowing other researchers to check the work.