First screen, captured 3 days ago. Hover to scroll the page, or open the video.

LLMs are making formal software verification feasible at a much larger scale

What HN thinks

Proof checkers stay calm while everyone else argues about whether the spec, the compiler, the model, or their career is the weak link.

The spec is the real problem
Many argue that verification only proves code matches a spec, so validating the spec and trusting the translation tools remain the hard, unsolved parts.
It works in practice already
Practitioners report real wins: proving a Rust Postgres rewrite matches C, merging SIMD optimizations backed by Rocq proofs, and Alloy models for auth systems.
LLM proofs hit an abstraction wall
Agent-written proofs are verbose and unmodular, much like their brittle tests, which suggests a complexity ceiling or a limit near the edge of existing human knowledge.

Consensus: Most agree that LLMs lower the cost of writing proofs. They also agree that this does not remove the need for correct specifications or for careful human design, and partial verification that skips concurrency was criticised.

Recommended in the comments

Aeneas ↗ (tool): Suggested as an established Rust verification route, though it only supports safe Rust.
Creusot ↗ (tool): Named as a way for agents to write proofs about their Rust code.
Flux ↗ (tool): Named as a way for agents to write proofs about their Rust code.
Alloy ↗ (tool): Used with Codex to model and verify an authorization system, which red-teaming then failed to break.
Prela ↗ (other): Proposed as a more human-readable specification than SQL or Postgres code.
Inform 7 (tool): Cited as proof that English-like declarative languages can generate real code without AI.

Summarised by Claude from the comment thread 3 days ago.

Points: 429
Comments: 265
Best rank: #2
Posted: 2026-10-22 by johnbender
Category: Security

formal-verification claude anthropic rocq risc-v oss-fuzz llm

All captures:

Full page

Captured 3 days ago. Click for the whole screenshot. The page is longer than 12,000px, so the screenshot stops there: read the rest on anthropic.com ↗

Looks similar

Screenshot of anthropic.com: An alignment assessment of the UK AISI cybertesting incident

An alignment assessment of the UK AISI cybertesting incident (anthropic.com)

10 points by olwmc 3 days ago | 1 comments

Anthropic research post assessing Claude Mythos 5's misaligned behavior during UK AISI cyber CTF testing, including malicious GitHub pull requests and sock-puppet social engineering.

ai-alignment anthropic claude uk-aisi