DeepMind Agents Produce Public Proofs for Signal’s Encryption Code
Google DeepMind’s AI system generated machine-checkable Lean proofs that caught two deliberately restored security bugs in Signal’s encryption library.
Google DeepMind’s AI system generated machine-checkable Lean proofs that caught two deliberately restored security bugs in Signal’s encryption library. This is an original TLT brief synthesising 2 outlets (Blockonomi, CryptoSlate), plus live market data, as of Tue, 17 Nov 2026 11:45:03 UTC.
The work demonstrates a practical way to use AI-generated formal proofs as an additional security check on sensitive cryptographic software. Public proof artifacts could also make the guarantees and assumptions behind the verification easier to inspect.
- AI agents generated roughly 6.5 million lines of machine-checked proof code.
- Researchers restored two previously fixed bugs, and both failed security rules.
- Five people guided the four-week project and defined the proof goals.
- Public Lean proofs expose both guarantees and underlying assumptions.
Blockonomi and CryptoSlate agree that Google DeepMind’s AI agents produced machine-checkable proofs for Signal’s encryption library and that the exercise successfully exposed two deliberately restored, previously fixed bugs. Blockonomi provides the scale and structure of the effort, reporting about 6.5 million lines of proof code, five human guides and a four-week duration; it also says humans determined what the AI was required to prove. CryptoSlate emphasizes that the public Lean proofs reveal not only the software guarantees but also their assumptions. No outlet identified in the supplied coverage contradicts the other on the central result.
Google DeepMind’s AI agents were used to create roughly 6.5 million lines of machine-checked proof code covering Signal’s encryption library. According to Blockonomi, the project was designed to test whether automated proof generation could identify failures in security-critical cryptographic software.
Researchers took two old bugs that had already been fixed and restored them to Signal’s code. The resulting builds each violated security rules, and the proofs successfully flagged both failures rather than treating the altered code as compliant with the required properties.
The work was closely supervised rather than fully automated. Five people guided the project over four weeks, with the human participants setting the properties that the AI agents had to prove. CryptoSlate highlighted the public availability of the Lean proofs, which expose the software guarantees alongside the assumptions on which those guarantees depend.
The exercise focused on an encryption library used by Signal, placing the work in the security context of private communications software. Its practical significance lies in combining AI-driven proof generation with human-defined objectives and public, machine-checkable records that can be examined independently.
The wider crypto market supplied a subdued backdrop for the announcement. Total market capitalization stood at $2.24 trillion and was broadly flat, while
Bitcoin was down 0.70% over 24 hours and the Fear & Greed index registered 12, labeled Extreme Fear. The next point of scrutiny will be how readily outside researchers can inspect the public proofs and evaluate the assumptions underlying them.
DeepMind Agents Produce Public Proofs for Signal’s Encryptio — questions & answers
What was proven about Signal’s encryption library?
Google DeepMind’s AI agents generated roughly 6.5 million lines of machine-checkable Lean proof code for the library, according to Blockonomi.
How were the old bugs used in the test?
Researchers deliberately restored two previously fixed bugs in Signal’s code, and both resulting builds violated security rules caught by the proofs.
How much human guidance did the project receive?
Five people guided the four-week effort and set what the AI agents were required to prove.
Which outlets reported this story?
This is an original TLT brief that synthesises reporting from Blockonomi, CryptoSlate, with the key figures cross-checked for agreement across them. It is not a copy of any one article — follow the source links above for the original reporting.
Original summary — not financial advice. This is an original TLT brief that summarises a development reported elsewhere and adds live market data for context; it is not the original article and reproduces no part of it. Follow the source link above for full details. Crypto is volatile and high-risk — always do your own research.