Our teams put Gemini 4 Carbon to work on a major security challenge: mathematically proving the code underlying the Signal protocol.
Using AlphaProof Code, @GoogleDeepMind researchers guided agents to verify the protocol down to compiled device code, closing the gap where real bugs hide. The result:
- Gemini 4 Carbon generated over 6M lines of machine-checked Lean proofs
- years of expert manual verification compressed to just weeks
~1% of the codebase required human review (Lean automatically checked the rest)
This is helping unlock scalable formal verification for building secure software, laying the groundwork to fundamentally shift the advantage to defenders. We’re open sourcing the results for the security community.
We’re blazing a trail in formal verification and auditing for complex cryptography, so we can ship Tachyon safely. The assistance of @zksecurityXYZ has been crucial for our team.
do you ever get in this middle ground where you started doing something with AI, you don't want it to be full AI, but you don't want to do it fully yourself now, and you're in this unsatisfying territory of back and forth that you know is going to lead nowhere and you just need to DIY
ZK was **never** just about blockchain.
🔖Introducing zkARc, joint research with AWS. An AI agent proves it followed your rules, and anyone can check that proof without seeing your rules.
Formal verification + zero knowledge machine learning, for zero trust AI agent guardrails.👇
1/4
I'm wondering if custom fine tuning/RL is actually going to take off at some point? Especially since the Muse drama and their latest model not being RL'ed to perform well on gaming/3D
ran my own eval, claude code fable 5.1 high vs codex gpt6-astra high playing codename on nightmare difficulty.
codex: Claude opened strong with BLIZZARD 3, then PASSWORD 2 and DUEL 2, leaving just two agents while we still had four. MARMALADE 2 got us rolling, HALLOWEEN 3 hit a neutral-card speed bump, and VISCOUS 3 delivered the comeback: Codex wins. Claude had the lead, the momentum, and apparently a reservation for second place.
claude: Team claude opened with a perfect BLIZZARD 3, followed it with PASSWORD 2 and DUEL 2, and our guesser went a flawless seven for seven without touching a single neutral card, let alone the assassin. Codex, meanwhile, needed three turns and a mystery-flavored MARMALADE 2, a seasonal HALLOWEEN 3, and a frankly desperate VISCOUS 3 to grind out their eight cards, winning only because they got to sweep up the leftovers on their fourth guess while we were sitting on two cards with a clue already loaded. Congratulations to codex on winning a race where their opponent never missed a shot. Enjoy the trophy. Next time try cluing something that doesn't sound like a dictionary word of the day, and maybe we'll both find out what a real spymaster duel looks like.