Google DeepMind AI agents prove key security properties of Signal's code
DeepMind says its AlphaProof Code agents proved key security properties of libsignal down to machine code. The proofs cover one-to-one chats only, and cost was not disclosed.
Key takeaways
- Google DeepMind says AI agents in its AlphaProof Code framework proved key security properties of libsignal at the machine-code level.
- The proofs cover the PQXDH handshake and Triple Ratchet, including forward secrecy, authentication and post-compromise security, mostly against classical rather than quantum attackers.
- Five researchers worked for four weeks; agent teams wrote about 6.5 million lines of Lean proof, according to the paper.
- No new flaws were found in the examined code, and the method caught two old, already fixed bugs that were reinserted as a test.
- The work covers one-to-one chats only, not group messaging, key generation or the full Signal app, and Google did not disclose its cost.
Google DeepMind says a team of AI agents has mathematically proved that the reference code behind the Signal encryption protocol meets several of its main security guarantees. The work, published on Nov. 10 in a technical paper alongside a framework DeepMind calls AlphaProof Code, checked libsignal as compiled machine code rather than a simplified model of it. DeepMind's researchers say that, to their knowledge, no one had previously proved protocol-level security properties for an unmodified, widely deployed implementation of a cryptographic protocol.
What did DeepMind prove?
Formal verification uses mathematics to show that software does what its specification says, ruling out every bug that would break the stated properties rather than hunting for bugs one at a time. According to the paper, the proofs cover Signal's post-quantum PQXDH handshake and its Triple Ratchet, and establish forward secrecy, authentication, post-compromise security and handshake secrecy. These properties are proved against conventional, or classical, attackers; only handshake secrecy is also shown against a future quantum attacker, in the paper's symbolic model. The team verified an unmodified libsignal release, linked into a small command-line test program and built with the standard Rust compiler for 64-bit Arm Linux.
The Signal protocol is used in the Signal app and, in variants, in WhatsApp, Meta's Messenger and Google Messages, the paper notes. Wired describes the result as further independent validation of the cryptography protecting Signal, Google Messages and WhatsApp.
How did the agents do it?
In DeepMind's setup, people decide what should be proved, and AI agents help turn the security goals into formal specifications in Lean, a language for machine-checked mathematics, with human experts approving each specification. Teams of prover agents then write the proofs, and a proof is accepted only if Lean's kernel checks it against a protected copy of the approved specification.
The paper says five researchers did the work over four weeks. Specifications were written with agents powered by Google's Gemini 4 Carbon and Anthropic's Claude Opus 5.5, while the proofs themselves came entirely from agent teams running on Gemini 4 Carbon; none of the authors wrote proof code. A typical proof campaign used 40 to 80 agents on a single machine for one to three days. The result is about 6.5 million lines of Lean proof, resting on about 26,000 lines of specifications and machine-code semantics that a reader has to trust.
For comparison, the paper points out that verifying the seL4 microkernel, one of the field's best-known successes, took twenty person-years.
Did it find any bugs?
No new flaws were found in the parts of the protocol that were examined, Wired reports. To test whether the method catches real mistakes, the researchers put two old libsignal bugs, which Signal had long since fixed, back into the code. In both cases they proved that the resulting builds broke the security properties, according to the paper.
Signal's chief technology officer, Ehren Kret, said in a statement quoted by Wired that DeepMind's approach allows a significant expansion on earlier verification work, which had focused on smaller portions of the code.
What are the limits?
The proofs cover less than the whole of Signal. The paper covers one-to-one sessions only and leaves group messaging for future work. It does not verify key generation, and it assumes that the low-level cryptographic building blocks behave as specified. It also does not cover multi-device setups, or prove the post-quantum healing that Signal's newest ratchet layer is designed to add. Wired notes that the work did not verify the whole Signal mobile app. DeepMind's researchers told Wired they had something of a head start from earlier formal verification work by the Signal Foundation.
Cost is an open question. According to Wired, DeepMind did not charge for the work, which it ran as research. Google would not share the cost of the project or say whether a commercial verification product or service is planned, according to Wired, which adds that it is unclear how practical AI-aided formal verification will be at scale. "We are very excited by this result, but there is a long distance to go before we are able to scale this to all the software of the world, and we want to do that," DeepMind research scientist Swarat Chaudhuri told Wired.
DeepMind is publishing its specifications and proofs so that others can check them. The paper calls the work a milestone towards Signal Shot, a challenge that Signal, the Lean FRO and the Beneficial AI Foundation launched in April to formally verify the Signal protocol and app in Lean.
Questions readers are asking
Does this mean Signal is proven secure?
No. The paper proves specific properties for one-to-one sessions in libsignal, assuming the underlying cryptographic primitives work as specified. Group messaging, key generation and the Signal apps were not verified.
Which AI models were used?
According to the paper, specifications were written with agents powered by Gemini 4 Carbon and Claude Opus 5.5, and the proofs were produced by agent teams running on Gemini 4 Carbon.
Did the agents find new vulnerabilities?
Wired reports that no new vulnerabilities were found in the parts examined. When two previously fixed bugs were reinserted, the researchers proved the resulting code broke the security properties.
Can others check the work?
Yes. DeepMind says it is releasing its Lean specifications and proofs publicly.
Sources
- #google-deepmind
- #signal
- #formal-verification
- #ai-agents
- #encryption
- #lean
Get the day's top AI & tech stories on WhatsApp: Follow the NextFish channel
Get stories like this every morning.
Related
Agents3 min
OpenAI launches ChatGPT Wallet so its dots agents can buy for users
The payment feature starts with US ChatGPT Pro subscribers. Users approve each purchase with a biometric check, and agents get a payment token, not the real card number.
Agents3 min
Google opens Antigravity's /teamwork agents, claims math and code results
Google Research says its multi-agent system helped prove a 2013 conjecture and port large codebases to Rust, and offers 10,000 academics free access. Results are Google's own.
AnalysisAgents3 min
How big shopping and travel sites now write rules for AI agents
eBay, Target, Booking, Delta and Shopify have all added terms about AI agents. NextFish compared their terms with archived copies to see what changed and when.
Agents3 min
OpenAI's ChatGPT Wallet lets its dots agents buy things for users
Built with American Express, Mastercard and Visa, it starts with some US ChatGPT Pro users. Each purchase needs user approval, and OpenAI says agents never see card details.