DeepMind uses AlphaProof Code to formally verify Signal's protocol
Gemini Carbon agents produced over six million lines of Lean in four weeks and flagged two reintroduced bugs.
Stories per day · 30 days
2 RSS
Relatedlean18signal6alphaproof3autoformalization3cryptography3deepmind3openai3ai-agents2alphaproof-code2anthropic2claude2complexity-theory2gemini2libsignal2rocq2smt2theorem-proving2aegis1
57 stories, newest first · one card per story
Stories tagged #formal-verification
Google widens Gemini 4 Carbon access after Signal verification
Paying users get API access at $2 and $10 per million tokens until November 30; Carbon for cyber stays in Fairwind.
Paper sets out information-flow security proofs for quantum programs
Noninterference and simulation-based definitions are proved equivalent, with verification via quantum Hoare logic and self-conjugation.
Axiom Math says it formally verified the Stacks Project
Agents turned about 7,654 pages into roughly 16 million lines of Lean checked by a proof assistant.
1
Lasm targets formally verified web apps in Lean
The Hacker News post has 21 points and 3 comments; the feed gives no technical or release details.
FARMS maps computational mathematics into reusable, Lean-checked knowledge
Of 171 numerical linear algebra and PDE targets, 157 were accepted at a mean estimated cost of $0.85–2.29 each.
Ramanujan Challenge paper reports an AI pipeline for proof judging
The pipeline cut false-positive verdicts on natural-language proofs from 23% to 6% versus one-shot LLM judgment.
Anthropic hiring manager drops coding interviews as AI labs ship updates
The roundup also covers Genie 4's limited preview, a Lean formalization of CFSG, and disputed ETH claims.
EasyProver solves near-saturated Lean benchmarks and half of harder Aegis tasks
On a stratified Aegis sample of 2,000 statements, it proves 1,054 within a 30-minute budget.
OpenAI asks experts to check AI proofs in math, CS, and physics
OpenAI says models disproved the Exponential Time Hypothesis and proved local smoothing in every dimension; reviewers have not endorsed the claims.
2
OpenAI shares AI math, complexity and physics results with academics
Shepherds chose three release paths; Lean writeups for the math and ETH results are already public.
1
Simple Lean 4 harness h1 matches specialized prover workflows
With matched models, a single attempt plus verifier feedback solved eight open FormalConjectures problems and led three formal benchmarks.
Claude produced formal proofs for code from 123 OSS-Fuzz libraries
Rocq checks the proofs, which hold only for one architecture, single-threaded code, and inputs under 64 MB.
2
Anthropic says Claude proved properties in 123 open-source libraries
Proofs cover single-threaded RISC-V static binaries and inputs under 64 MB; a subset and a Claude skill are being published.
1
FVQKD machine-checks finite-key BB84 security in Rocq
The admit-free proof surfaces unstated side conditions, including a missing branch in the published argument.
PaperFormalizer formalizes research papers in Lean with the author in the loop
It runs in Claude Code or Codex, pauses on ambiguous claims, and emits an annotated paper plus linked Lean code.
AI agents discover A-greement, a faster geo-replicated consensus protocol
Across six 5- or 9-replica deployments, mean latency fell 12.2% on write-heavy and 33.3% on read-heavy workloads.
PerMIT adds machine-checked integrity controls for LLM agent memory
On Trojan Hippo, the authors report zero attack success when self-verification and lightweight verifiers confirm memory.
LLM2Cert synthesizes certificates checked by SMT solvers and Lean
The method verifies 81 of 105 DynaCertBench instances, 17.1 points above one-shot proposals without feedback.
RTL repair scores can miss whether models used the broken design
A new 459-case benchmark finds simulation and formal scoring can disagree, and specification-only baselines swing widely.
Paper proposes proof-carrying contracts for AI agent agreements
An Airlock blocks an operation unless the agent supplies a proof that agreed constraints hold in every modeled case.
doubleAI unveils Maxwell, a runtime model that rewrites running binaries
doubleAI says Maxwell proves each rewrite equivalent to the original and claims large speedups on GPUs and CPUs.
Blame-aware recomposition extends secure compilation proofs to low-level isolation
The technique allows compartments to observe each other's code and is applied to a tag-based reference monitor.
Mathematicians say OpenAI doesn't understand its own Navier-Stokes proof
Mathematicians dispute OpenAI's Navier-Stokes Lean proof, saying it differs from the cited approach and agent credit is unclear.
4