DeepMind's AlphaProof Code machine-verifies security properties of Signal's crypto code
Fluffy_Department_21 · reddit · 2026-11-18
Google DeepMind says its AlphaProof Code system helped verify specific security properties of the compiled code behind Signal's core cryptographic protocol.
- Human experts reviewed mathematical specs, AI agents generated proofs, and the Lean proof checker machine-checked them
- Key distinction: this verifies implementation, not protocol design—a secure design can still be implemented incorrectly
- DeepMind also reintroduced two previously fixed bugs into an isolated copy; the system produced machine-checked proofs those versions were insecure
- Limitations: results apply only to the code and properties checked, and depend on the specs and assumptions being right
The post also invites discussion on where software security practitioners would apply AI-driven verification first.
More from Safety
- Security researcher pushes "Agentic Offensive Frameworks" term, warns against overusing "ARTEX" — cyb3rops · 2026-11-18
- AWS lays out a three-phase responsible AI roadmap for agentic AI platforms — AWS ML Blog · 2026-11-18
- Redwood AI launches podcast with first episode on OpenAI's monitoring stack — jammastergirish · 2026-11-18
- MIRI's Nate Soares on why we can't just unplug AI: three core arguments — jammastergirish · 2026-11-18
- Zhipu stock down ~33% in 4 days as 'darkest moment for Chinese AI' case spreads — Gregory_C_Allen · 2026-11-18
- Google finally lets you opt out of Gemini AI training without losing chat history — Ars Technica AI · 2026-11-18