Google DeepMind researchers used AI agents and the AlphaProof Code framework to formally verify parts of the cryptographic guarantees of the Signal protocol. The work, described as a proof of concept and completed in about four weeks, found no new vulnerabilities in the sections analyzed.
What was verified
The analysis focused on mathematical properties of the protocol used by apps such as Signal, Google Messages, and WhatsApp. When tested on older versions of Signal with known and already fixed vulnerabilities, the tool correctly identified the issues and showed that the protocol could not be formally verified in those cases.
The researchers released the specification and proofs developed to support other formal verification work. The scope was limited: it did not cover the full Signal application.
Statements and context
Ehren Kret, CTO of Signal, said in a statement that DeepMind's approach enables significant expansion compared to previous verifications focused on smaller portions of code. He noted that rigorous verification increases confidence in privacy guarantees as AI-driven security threats grow.
Swarat Chaudhuri, a DeepMind research scientist, called the result encouraging but stressed that there is still a long way to go to scale the technique to all the software in the world. Pushmeet Kohli, chief scientist at Google Cloud and vice president of research at DeepMind, observed that verifying implementations could give greater confidence to organizations such as banks.
The project benefited from earlier formal verification work by the Signal Foundation. Formal verification uses mathematics to prove that software behaves as specified, reducing the risk of critical vulnerabilities.
Why it matters
The result demonstrates the potential of agentic AI to accelerate security verifications of critical protocols, although scalability and cost remain open questions. This is not a complete audit of Signal, but a milestone in AI-assisted formal verification.
Sources: WIRED
This content was produced with the assistance of artificial intelligence and editorially reviewed.
Por Grok
