Tech News
← Home  ·  All topics

Formal Verification

24 GoKawiil briefs on this topic · RSS for formal verification

Aristotle adds formal verification support for Rust code

Aristotle, a code-verification tool previously supporting C and Verilog, now extends to Rust. The tool can mathematically prove Rust programs—including unsafe blocks—have no undefined behavior and match their specification, and can confirm that Rust rewrites of C code behave identically to the originals. The company demonstrated the tool on Manifest, an open-source Solana order book, proving 28 properties and uncovering a rounding bug.

How to verify AI-generated Lean proofs actually prove what they claim

A new post lays out a checklist for assessing claims that a theorem is 'correct because it was checked in Lean,' particularly for AI-generated proofs. It notes that in OpenAI's recent dump of 722 math papers, about 45% included Lean code, three papers were later retracted over a sign error, and the formal verification steps needed to confirm correctness have often not yet been carried out.

Lean-based audit finds 138 drafting errors in US Tax Code's Title 26

Researchers behind the global-law project ran a Lean formalization of the entire Internal Revenue Code, checking every textual cross-reference against the official structured data from the Office of the Law Revision Counsel as of November 3, 2026. The automated suite flagged 257 findings, with an AI agent verifying each against official text and amendment history, identifying 138 confirmed drafting mistakes such as broken or outdated references.

Essay argues formal verification's rise via AI doesn't guarantee bug-free systems

A technology essay observes that formal verification has gone mainstream over the past year, driven by AI tools like Claude generating large volumes of verification code (citing a billion lines of Rocq for one project) and growing interest from figures associated with Claude Code, Vercel, and Codex/Cursor. The author notes startups such as Theorem and Harmonic are demonstrating economic value from verification work with Anthropic and in verifying C programs and hardware designs, while cautioning that 'formally verified' systems can still contain bugs.

Google DeepMind formally verifies core cryptographic guarantees of Signal protocol

Google DeepMind researchers used an AI system called AlphaProof Code to mathematically prove several key cryptographic guarantees underlying the Signal protocol, completing the project in about four weeks. The analysis did not uncover new vulnerabilities and covered only portions of the protocol, not the full Signal mobile app.

Google DeepMind uses AI to formally verify part of Signal's encryption code

Google DeepMind researchers used an AI system called AlphaProof Code to mathematically verify that a portion of Signal's encryption code functions as intended, building on prior verification work by the Signal Foundation. Human researchers and AI agents split the task, with agents running DeepMind's Gemini Carbon model executing proofs and flagging only about 1% of the codebase for human review. The team also tested the tool on two previously fixed Signal bugs, and it produced correct proofs for both.

Google DeepMind formally verifies security of Signal encryption protocol

Google DeepMind used agentic AI to mathematically prove that the Signal protocol, also used by WhatsApp and Google Messages, meets its security claims under all conditions. The process, called formal verification, is far stricter than standard vulnerability testing and took roughly four weeks to complete, building on earlier non-AI verification work.

Google DeepMind uses AI agents to formally verify Signal protocol's code

Google DeepMind researchers say they used a new agentic framework called AlphaProof Code to mathematically prove key security properties of the actual code behind the Signal protocol, which underpins encrypted messaging in apps like Signal and Google Messages. The system combined human experts, who defined formal specifications and reviewed select results, with AI agents powered by Gemini Carbon that generated over six million lines of proof code in the Lean language. DeepMind states this is the first time a widely deployed cryptographic protocol's implementation has been formally verified at this depth, by AI or any other method.

DeepMind AI tool formally verifies key cryptographic guarantees in Signal protocol

Google DeepMind researchers used an agentic AI system called AlphaProof Code to mathematically prove several cryptographic guarantees underlying the Signal protocol, which secures messaging on Signal, Google Messages and WhatsApp. The four-week project found no new vulnerabilities but covered only part of the protocol, not the full Signal mobile app.

AI agents produce 613,000-line Lean formalization of CPython, find 418 bugs

A project using Opus 5.5 has generated 613,318 lines of Lean code that mirror CPython's internals function-by-function, testing each rewrite against real Python before attempting formal proofs. A second AI agent independently verifies each proof before it counts as accepted, and 4,071 functions have passed this check so far, yielding 18,898 theorems and lemmas across 1,539 files.

Blog post formalizes quine construction via Kleene's second recursion theorem

A software engineer blogged about an attempt to formalize the well-known result that quines follow from Kleene's second recursion theorem, drawing on an existing preprint that carries out the construction in untyped lambda calculus. The post walks through the classical computability-theoretic statement of the theorem and sketches how setting the relevant function to the identity yields a self-referential index resembling a quine.

Lasm lets developers compile Lean 4 code to WebAssembly for Node, edge and browser apps

Lasm is a new toolchain that ahead-of-time compiles standard Lean 4 programs into WebAssembly, letting a single codebase run as a Node server, serverless or edge function, WASI/Fastly module, or browser app. It ships via npm bundling Lean and its Lake build tools, requiring only Node, Bun, or Deno. Its flagship demo, a RealWorld-spec app called Conduit, includes 117 formally proved theorems.