Google DeepMind confirms the reliability of its AlphaProof Code tool with a test using planted bugs, plans to extend it to other critical systems
During a test of the AlphaProof Code tool, Google DeepMind deliberately reintroduced two previously fixed bugs into the Signal protocol's code — the system detected them immediately. The company plans to extend the method to power grids and medical devices. The verification of the Signal protocol itself took four weeks.
Google DeepMind, while verifying the reliability of its AlphaProof Code tool, deliberately reintroduced two previously fixed bugs into an isolated version of the Signal protocol's code. In both cases, the system generated a machine-checked proof that this version of the code was insecure, correctly distinguishing between working and buggy code. According to Signal Foundation CTO Ehren Kret, this approach strengthens confidence in the protection of users' private data, because as AI threats grow, it allows for more thorough verification of the protocol's implementation than previous methods. Google DeepMind says it plans to extend the formal verification method to other critical systems, such as power grids and medical devices, for example through the CodeMender tool.
This development follows an earlier report that Google DeepMind, using the AlphaProof Code agentic framework, formally verified the security of the Signal protocol's implementation down to the level of compiled code, not just its design. Experts established formal security specifications, and AI agents built on the Gemini Carbon model then generated more than six million lines of proof in the Lean language, with humans manually checking only about one percent of the code. The entire process took four weeks, whereas comparable manual verification, according to the company, has historically taken years.
The verification found no new vulnerabilities in the verified parts of the Signal protocol. For older versions with already known and fixed bugs, the tool correctly identified that the protocol could not be formally verified. The company plans to publish the specification and proof for further independent verification. According to Wired and Ars Technica, it remains an open question whether this approach can be scaled to a broader range of software and at what cost — Google DeepMind has not disclosed the project's costs and declined to comment on any potential commercial product.
Why it matters
A test using deliberately planted bugs shows that the AlphaProof Code tool can reliably distinguish between secure and buggy code, which increases the credibility of the method beyond a one-off demonstration on the Signal protocol. Plans to extend it to power grids and medical devices signal that AI-driven formal verification could eventually also serve to verify the security of critical infrastructure, where bugs carry high risk. Sources consistently point out, however, that it remains unclear how scalable and cost-accessible the method is outside of a free research project.
What was added since the original report
Verified updates
-
Two deliberately planted bugs in the code were caught immediately during verification; DeepMind plans to extend the AlphaProof Code method to power grids and medical devices; Signal's CTO confirmed that the approach strengthens confidence in data privacy protection
- Two deliberately planted bugs in the code were caught immediately during verification
- DeepMind plans to extend the AlphaProof Code method to power grids and medical devices
- Signal's CTO confirmed that the approach strengthens confidence in data privacy protection
-
AI agents generated more than 6 million lines of Lean code as part of the verification; Humans manually reviewed only 1% of the codebase; Verification took place at the level of the compiled code implementation, not just at the design level; The framework combines human expertise with the autonomous work of AI agents; This is the first time a cryptographic protocol has been formally verified at the implementation level by AI
- AI agents generated more than 6 million lines of Lean code as part of the verification
- Humans manually reviewed only 1% of the codebase
- Verification took place at the level of the compiled code implementation, not just at the design level
- The framework combines human expertise with the autonomous work of AI agents
- This is the first time a cryptographic protocol has been formally verified at the implementation level by AI
Two audiences, two different impacts
What this means
For individuals
Users of Signal and Google Messages have further independent confirmation that the encryption of their messages works according to the specified security guarantees, down to the level of the code that actually runs.
More practical updates →For a business
Companies operating critical infrastructure, such as energy utilities, healthcare facilities, or banks, are getting a signal that AI-driven formal verification of code could in the future become an available way to demonstrate security guarantees; for now, however, it remains a free research project with no published pricing or commercial offering.
Risks and complianceCheck the original
Event sources
confirmed by 3 independent sources · 4 publishers, 3 independent. We count feeds from the same owner only once.