CONFIRMED
Google DeepMind announced AlphaProof Code on November 10, an artificial intelligence agent framework that helped researchers mathematically prove security properties of Signal's cryptographic implementation. The important consequence is not that a messaging application has suddenly become invulnerable. It is the prospect of bringing formal verification closer to the software actually running on devices, rather than checking only the theoretical design of a protocol.
From security goals to running code
Formal verification turns desired guarantees into mathematical statements and demonstrates that an implementation satisfies them under specified assumptions. Traditional tests look for failures in selected situations. A formal proof establishes a property across the cases covered by its model. That distinction matters when an unobtrusive implementation mistake could compromise private conversations or critical infrastructure.
DeepMind says human specialists reviewed every specification before agents powered by Gemini Carbon built proofs in Lean, a language and tool for mathematical verification. The work reached properties of compiled code, including message protections before and after a temporary device compromise. Lean checks the proofs, so the outcome does not rest solely on trusting an answer generated by another model.
Four weeks, with experts still involved
The project took four weeks. Independent reporting by WIRED notes that earlier Signal verification research gave the team a head start. It also stresses that the researchers did not verify the entire mobile application. DeepMind describes the work as unprecedented at this depth, a claim by the laboratory that should not be mistaken for saying nobody had previously analyzed the protocol.
To check whether the framework could distinguish correct implementations from flawed ones, researchers reintroduced two real, previously fixed bugs into an isolated version. According to the announcement, the system produced checked proofs of insecurity in both cases. This does not disclose a new vulnerability in today's Signal. It is a control showing that the approach can reject insecure implementations rather than simply endorse versions already considered safe.
A practical lens for smaller businesses
For small and medium businesses in Brazil, the immediate value is mainly a better standard for evaluating software suppliers. A platform claiming to use AI for security should explain exactly which properties it verifies, which components are covered and which assumptions apply. Checking cryptography does not establish that permissions, integrations, backups or automated customer support behave correctly.
Neither source establishes commercial pricing or ordinary team access to AlphaProof Code. WIRED reports that Google did not disclose the research cost or confirm plans for a commercial product. Releasing verification artifacts lets specialists inspect the work, but it does not automatically make the framework a tool available through an existing Gemini subscription.
Our editorial assessment is that the most promising contribution is the combination of human chosen goals, agents building proofs and a checker separate from the generated prose. This could reduce specialized effort, but the study does not establish broadly applicable business savings. The advance deserves attention precisely because it replaces generic confidence in AI with narrow, inspectable guarantees, without eliminating engineering work, human review or assessment of risks outside the proof's scope.
Sources
Google DeepMind: Introducing AlphaProof Code; WIRED: A New AI Tool From Google DeepMind Proved the Signal Protocol Is Secure.
