
@chbpap presents our deepprove paper at ACM CCS in the Hague on wednesday, the one @AnthropicAI called the current peer-reviewed state of the art. If you’re there, go find him. what do you want to ask him?

professor of computer science at @Yale | co-director of @YaleACL | chief scientist at @lagrangedev | researching and teaching cryptography and security

@chbpap presents our deepprove paper at ACM CCS in the Hague on wednesday, the one @AnthropicAI called the current peer-reviewed state of the art. If you’re there, go find him. what do you want to ask him?

Yale @YaleACL @YaleCSDept @YaleEngineering researchers are in the Netherlands this week presenting 9 papers at ACM CCS 2026!
🎉 Yale at ACM CCS 2026! Excited to share that 9 papers co-authored by Yale researchers will be presented at #ACMCCS2026 @acm_ccs! Check out our papers below! 🧵👇

DeepProve, built by @lagrangedev, is cited in new research from @AnthropicAI as 'the current peer-reviewed state of the art' for provable inference.
The Frontier Red Team at @AnthropicAI just made the case that you should be able to check which AI model answered you, without ever seeing its weights. Their research names DeepProve “the current peer-reviewed state of the art.”

We’re watching crypto “grow-up” in real time. Adoption of ZK has now reached the largest frontier labs in AI. Excited to see @AnthropicAI recognize @lagrangedev’s DeepProve the “current peer-reviewed state of the art” for verifiable AI.
New Anthropic research: progress on provable inference for large language models. Here, we introduce FlashProve. Building on established cryptographic techniques, FlashProve allows developers to verify whether a given response came from a given AI model. It’s substantially more efficient than prior open-source methods for doing so. These proofs have a range of applications. They can, for instance, make it harder for attackers to hijack a chat, or enable safety evaluators to verify that the model they’re testing is the same one being served to customers. For our full write-up, see here: anthropic.com/research/prova… Further progress in provable inference will likely be an important part of improving trust in AI deployment. To that end, we're open-sourcing FlashProve’s code, here: github.com/anthropics/FlashP… We’re grateful to our collaborators: Alessandro Chiesa, @GiacomoFenzi, @gwberg95, @miranda__christ, Rahul Dalal (rdalal.github.io) and Ben Mayer (@transmissions11).

I am happy to share FlashProve: our performant open-source prototype for provable inference. We believe that this has significant implications for security, pacing AI development, integrity guarantees, and more. FlashProve allows a provider to commit to model weights, and then prove that replies originate from the committed model. The proof leaks no information about the weights, and the verifier can efficiently verify it on consumer hardware. Our library is performant, proving a 400B parameter MoE model at 100 tok/s. We have open-sourced the associated code and Lean formalization. We hope that it catalyzes further progress in verification of inference, and look forward to seeing it.
New Anthropic research: progress on provable inference for large language models. Here, we introduce FlashProve. Building on established cryptographic techniques, FlashProve allows developers to verify whether a given response came from a given AI model. It’s substantially more efficient than prior open-source methods for doing so. These proofs have a range of applications. They can, for instance, make it harder for attackers to hijack a chat, or enable safety evaluators to verify that the model they’re testing is the same one being served to customers. For our full write-up, see here: anthropic.com/research/prova… Further progress in provable inference will likely be an important part of improving trust in AI deployment. To that end, we're open-sourcing FlashProve’s code, here: github.com/anthropics/FlashP… We’re grateful to our collaborators: Alessandro Chiesa, @GiacomoFenzi, @gwberg95, @miranda__christ, Rahul Dalal (rdalal.github.io) and Ben Mayer (@transmissions11).

Incredible to see @AnthropicAI refer to @lagrangedev’s DeepProve as the “the current peer-reviewed state-of-the-art” for verifiable inference. The frontier of ZK, verifiable and private inference is moving faster than ever!

here are my thoughts about OpenAI's amazing Mat Mul result. comments very welcome.

This is why we work on Maverick eprint.iacr.org/2026/1955. It’s a lightweight solution that doesn’t involve ZKP or FHE or MPC. While TEEs are also lightweight, there is a cost factor: you have to put entire workflow in some high-end Nvidia GPUs with TEEs, which can be an overkill.

This is a spectacular result, saying the NP complete problem over n variables can be solved in much less than 2^n time. I love it, always thought that SAT shouldn’t be as hard as 2^n. It doesn’t mean P=NP, though.
Here's a direct link to OpenAI's PDF claiming a refutation of ETH. In particular, it's claimed that 3SAT on n variables is in about exp(n/(log n)^{1/2+o(1)}) time github.com/openai/math/blob/…

Continuing the work started at @lagrangedev, I am excited to announce a @Yale @Penn collaboration on a black-box compiler from static to dynamic SNARKs, yielding post-quantum dynamic SNARKs. Work led by amazing @BawejaAnubhav from @Penn! eprint.iacr.org/2026/2445
Tweak an AI model's weights and the proof of what it computed starts over from zero. @chbpap shows how dynamic zk-SNARKs update the proof when the model does.

Impressive!
Dolev–Strong is arguably the most well-known consensus protocol, yet in about 45 years no deterministic protocol improved on its O(n^3) communication. With @komasito55 & Ling Ren, we get O(n^2 log^4 n) for Byzantine Broadcast, t<n faults, O(t) rounds: eprint.iacr.org/2026/2504

This podcast from @a16zcrypto includes an aspect that hasn't been stressed enough. In the end, AI cryptanalysis is a great thing for cryptography as it helps us gain confidence in new assumptions, key sizes (e.g. for new PQ schemes) and more. youtube.com/watch?v=k_CoOoDs…

The what now?! arxiv.org/abs/2610.06783v1

Our @YaleACL seminar this semester will mostly feature talks from distinguished colleagues on cryptography and AI. yacl.cs.yale.edu/seminar/

Formal verification of functional correctness is quickly becoming the bare minimum standard. It will soon be irresponsible for any enterprise to release code which is not formally verified. AI coding agents can generate not only the code, but also a spec with a proof that the code satisfies the spec. Or put simply, if it is so smart to generate the code, it should also be smart enough to tell me what it did and give me all the verifiable evidence that it did it. This used to be hard, but it is not anymore. The spec, any spec, forces the agent to reflect on its code symbolically and prove some properties about it. That will make it write better code and solve more challenges. Same like humans. The spec can and should include the known security attack vectors, for starters. There will always be unknown unknowns, but known unknows are not an option anymore, as they will be instantly exploited. Generating formally verified code is possible now. The agents can do it with the adequate harness and proving infra, at no extra effort from the devs.

7/ For more information on the technical aspects of the @lagrangedev DeepProve project, see our paper eprint.iacr.org/2026/1112.