Google DeepMind verifies Signal's cryptographic core in four weeks
DeepMind's AlphaProof Code, powered by Gemini Carbon agents, formally verified the cryptographic core of the Signal messaging app in four weeks, work that typically takes years.
Google DeepMind has used its AlphaProof Code system to formally verify the cryptographic core of Signal, the encrypted messaging app, in roughly four weeks. Conventional verification work of this kind usually takes years, which makes the speed of the project the central point of the report.
According to the reporting, AI agents powered by Gemini Carbon wrote the proof code. The agents produced more than six million lines of proof code. About one percent of that output was flagged for human review, meaning that the large majority of the generated material moved through the process without manual intervention, while the flagged portion was checked by people.
The method is designed to mathematically confirm that code behaves correctly, rather than relying only on testing to find errors. Formal verification proves properties of a program against a precise specification, and it is widely regarded as one of the strongest assurances available for security-critical software. The approach is labour-intensive when done by hand, which is why the reported timeframe stands out.
DeepMind plans to extend the approach to critical infrastructure. The systems named as targets are power grids and medical devices, where software faults can have direct physical and safety consequences. The reporting frames the Signal project as a demonstration that AI-assisted proof generation could make such assurance practical at a much larger scale.
The available account is limited. It does not give details of specific verification results, nor does it set out the exact specification that the code was checked against. It also rests on a single report, so the findings have not been independently reviewed in the material available here. What the report does establish is the scale of the effort, the division of labour between AI agents and human reviewers, and DeepMind's stated intention to apply the method beyond messaging software.
Source: the-decoder.com
17
