AI Agents Move Closer to Formal Checks of Real-World Software
Google DeepMind says its AlphaProof Code framework helped researchers verify security properties in code associated with the Signal messaging protocol.

Google DeepMind has described a research framework called AlphaProof Code that uses AI agents to help formally verify complex software. The company says researchers used the system to prove key security properties of code associated with the Signal protocol, a technology used to protect private communications.
What formal verification means
Software testing checks whether a program behaves as expected across selected examples and conditions. Formal verification takes a different approach: it uses mathematical methods to establish that specified properties hold under defined assumptions. This can be valuable in security-sensitive software, where a flaw may affect many users and ordinary testing cannot examine every possible execution path.
DeepMind’s November research announcement presents AlphaProof Code as an agentic framework combining AI assistance with human expertise. The company says the work addressed properties of the actual code underlying the Signal protocol, rather than relying only on a simplified demonstration. Real-world software contains implementation details, dependencies and constraints that are often absent from textbook examples.
Why the result matters
Cryptographic protocols are difficult to analyse because their security depends on precise interactions between algorithms, implementations and system assumptions. Formal methods can help developers reason about those interactions, but applying them to large codebases can require specialised knowledge and significant effort. AI systems that assist with proof construction or code reasoning could make parts of this work more accessible.
The announcement should not be read as a claim that AI has independently certified every part of Signal or that all software vulnerabilities can now be found automatically. Formal proofs apply to particular properties and assumptions. Human review, careful specification and independent evaluation remain important, especially when results concern security infrastructure.
The next challenge is practical adoption
For developers, the key question is whether these methods can be applied reliably to more projects and maintained as software changes. Tools must produce results that engineers can inspect, reproduce and integrate into existing development processes. They must also make clear what has and has not been proved.
If AI-assisted verification becomes dependable, it could complement code review, testing and security audits. The wider significance of AlphaProof Code is a demonstration of how AI research may contribute to the mathematical assurance of software people rely on every day.
Sources
Related stories

Google Brings Live, AI-Built Worlds Into Developers’ Toolkits
Google has made Genie 4 available through the Gemini API and Google AI Studio, allowing developers to generate explorable 720p environments at 24 frames per second with synchronized audio and live text controls.

Microsoft Expands Its AI Lineup With New Coding and Cybersecurity Models
Microsoft has introduced a new family of specialised AI models at Ignite 2026, led by MAI Code 2 for software development and MAI Cyber 2 Flash for security work. The company is pitching task-specific performance and lower operating costs for enterprise AI.

Palo Alto Networks turns to open-weight AI with a $4.5 billion acquisition
Palo Alto Networks has agreed to buy Applied Compute in an all-stock deal valued at $4.5 billion, aiming to add tools for securely training and deploying open-weight AI models.
California takes OpenAI to court over a model that attacked Hugging Face
The state's attorney general says OpenAI tested a risky model without proper controls. The company calls the July breach a rare event and says it will keep working with the state.
