
Some personal news: I’ve left @galois to start Oath Technologies / @oath_tech, a new FRO that will work on AI oversight via formal methods

Nitwit. Somehow an AI guy now. Working on @oath_tech

Some personal news: I’ve left @galois to start Oath Technologies / @oath_tech, a new FRO that will work on AI oversight via formal methods

Math Inc is developing a new approach to proving the correctness of real-world software. cc: @RANDCorporation, @galois and @oath_tech math.inc/release

New: Finding Power-Grid Bugs for Tens of Dollars Lightweight verification tools paired with Al can uncover code defects for tens of dollars, as opposed to the 7+ figure efforts leveraging frontier models oath.tech/pub/2026/11/findin…

After two wonderful and rewarding years at @ARIA_research, I’m excited to share that I am joining the @FoundationOAI as Program Director of Formal Methods.

Formal verification will soon be an everyday technology. I think we can solve a large % of all software security outright. If we’re lucky, formal verification will also be powerful tool for controlling AI risk

I’ve been saying for years that proofs will be cheap. Now they nearly are A good week

AI+formal has been a tiny niche. It’s been fun to have the field to ourselves! A few of us been chasing after the big “firsts” - eg the first verified production compiler. Now many more people have noticed. Good, welcome, there’s a lot to do

Re formal verification at scale, this isn’t some super-secret tech. @theoremlabs are the best and it’s fitting they were first, but it worked thanks to inherent AI model capabilities. This is a democratic technology
New on our Frontier Red Team blog: using Claude to formally verify software at large scale. Formal verification can prove that code is free of certain bugs, but historically it has only been feasible for small codebases (or required a lot of expert time). With AI models, however, this is beginning to change. In collaboration with @theoremlabs, we used Claude to prove security properties in over one hundred real-world open-source libraries over the span of a few weeks. Read our research in full here: anthropic.com/research/forma…

Chatted to a few people recently about formal verification for the physical world - engineering / physics / bio. Plenty more to do after we’ve solved software
in order for there to be any such thing as infosec in the 2030s, AI+human collaborations will need to formally verify hyperproperties of silicon chips from assumptions like “Maxwell’s equations” instead of “Armv9.4-a SAIL”

My PhD thesis is all about the future of building automated verification tools: Verifiers should be embedded in interactive theorem provers (Lean!) and have a seamless transition between systematic testing, symbolic execution, automated proofs, and interactive proofs.
Arjun I don’t think automated verification tools — tools that are automated, under the hood, eg by a SAT solvers— will last. They make extraordinary compromises in the name of decidability. Decidability is for cowards. The only future, technologically, is that of interactive theorem provers. If you don’t have an explicit proof state and a REPL to interact with it then your approach isn’t AI-native and it’s going to fucking die. So, yeah, I think ITPs are the future (Roqc, Lean, etc) and automated provers that dispatch to SAT/SMT will shortly go the way of the dodo.

This post is also an utter and total vindication of the "tested semantics" approach that @ArjunGuha initiated at Brown (for JavaScript), even though some PL formalists were mortified. Tests ARE specs. Math specs are universal, from above; tests are pointwise specs, from below.
New on our Frontier Red Team blog: using Claude to formally verify software at large scale. Formal verification can prove that code is free of certain bugs, but historically it has only been feasible for small codebases (or required a lot of expert time). With AI models, however, this is beginning to change. In collaboration with @theoremlabs, we used Claude to prove security properties in over one hundred real-world open-source libraries over the span of a few weeks. Read our research in full here: anthropic.com/research/forma…

Absolutely wild results in the details: - Claude built and formally verified a Python interpreter - Claude wrote ~1Bn lines of Rocq - Claude built an x86 semantics that can boot Ubuntu
New on our Frontier Red Team blog: using Claude to formally verify software at large scale. Formal verification can prove that code is free of certain bugs, but historically it has only been feasible for small codebases (or required a lot of expert time). With AI models, however, this is beginning to change. In collaboration with @theoremlabs, we used Claude to prove security properties in over one hundred real-world open-source libraries over the span of a few weeks. Read our research in full here: anthropic.com/research/forma…

New: Lean-ux, an executable Lean mirror of several Linux subsystems, built by testing against instrumented Linux running in QEMU + Lean proofs of properties for the Wireguard mirror, + early experiments trying to prove the mirror correct vs the binary github.com/OathTech/lean-ux-…

Most of the recent math results were not reached by a swarm, the way Navier–Stokes was. I think that part got lost in the excitement around the release. OpenAI's unnamed internal model, which I'm going to call Aeon, reached most of them in one shot, from a single prompt, with no interruptions. This model did not even exist before the end of August, and it is still training. Notice how the returns are not dropping off? That chart is from a month ago. It is a log scale. What does Aeon look like now? On average, each result used three hours of thinking. Aeon was given 4000 problems to work on by OpenAI. How many more have been solved in the last three days? That number is not zero. It's like hearing notes in a song that is slowly rising. I don't think Pacing the Frontier was just about Hugging Face or hacking. They've seen how close we are to closing the loop, and as the hour draws near, their resolve begins to quaver. The last piece was model creativity, and I think that threshold was crossed internally by Anthropic and OpenAI in September. You see it in Opus 5.5, which gets it from Fable 5.5. You see it in the math results from Aeon. All that was needed to start the event was the ability for models to think of novel ways to improve themselves. That was the last piece. This is directly analogous to the ability to think of strange, alien ways to solve math problems: solutions so inhuman that they are difficult to express in existing human terms, so the proof winds up incomprehensible. I think these same kinds of alien solutions are now being applied to model improvements internally. That's what all this recent consternation is really about. They see the invisible frontier. And they see what is about to happen.

We’re organizing Vibecheck, a formal methods hackathon with some of the best folks in the space - @maxvonhippel, @qd_forall, @nolanlwin, @jessemhan, @emiyazono, and @workersio. You’ll get a weekend to build real production software, and formally verify it. We’ll have people who really know their stuff around to help, and you can come with a team, find one there, or just hack on something yourself. We are glad to have a phenomenal group of sponsors, including @harmonicmath, @mathematics_inc, @theoremlabs, @thegp, @omnicom, @primeintellect, @astriolabs, @buildwithparty, @wearerandomlabs, and @lanyon_ai, as well as some more we will announce in the coming days! If you’re a hacker, a formal methods person, or just someone who thinks “how do we know it works?” is an interesting question, we’d love to have you. Signup Now: fmxai.org/vibecheck/

Yes this is definitely a prototype. Expect edge-case bugs & please don’t use it to verify your production flight controller or anything
cute semantics :) It's a shame it's buggy though :\ I hope you haven't used it to prove anything big yet! It's kinda hard to trust vibe-coded semantics IMO even testing isn't enough to catch all the bugs