Research

Advancing cryptography for AI agent verification.

Every agent action receipt ICME Labs issues rests on research our own team published. Each paper below advances one or more of the five requirements to meet before delegating high-stakes work to AI agents: a check that runs before the agent acts, a verdict that is the same every time with no model to hallucinate it, proof an outside party can verify independently, proof that travels beyond the platform that made it, and verification that exposes neither the rules nor the data.

Forthcoming with Amazon Web Services researchers Zero-knowledge receipts for agent guardrails
Advances: check before the agent acts, same verdict every time, outsiders can verify, proof travels, rules and data stay hidden

zkARc: Succinctly Verifiable Proofs for Neurosymbolic Natural Language Formalization and Verification

Alberto Centelles, Wyatt Benno, Benjamin Kiesl-Reiter (AWS), Aman Goel (AWS)

Today, the only evidence that a policy was enforced on an agent action is a log written by the same system that ran the agent, so a counterparty must trust the operator, re-run the check, or read the confidential policy. zkARc specifies a receipt that removes all three. Each stage of the check, from translating a plain-English policy into formal logic to the solver verdict, gets its own zero-knowledge proof, bound together so none can be swapped. A verifier confirms the check ran on the exact action received, against the authorized policy, before release, while the policy, model weights, and inputs stay hidden. Co-authored with the Amazon Web Services team behind Automated Reasoning checks. Preprint available on request.

arXiv, 2026 Verifiable inference
Advances: outsiders can verify, proof travels, rules and data stay hidden

Jolt Atlas: Verifiable Inference via Lookup Arguments in Zero Knowledge

Wyatt Benno, Alberto Centelles, Antoine Douchet, Khalil Gibran

When a language model is part of a compliance check, the buyer needs proof the model actually ran on the input in question. Jolt Atlas produces that proof: a committed neural network, a given input, a given output, and a succinct cryptographic argument that the three belong together. The model's weights and intermediate values stay hidden, so a vendor can prove correct execution without disclosing the model. This is the zero-knowledge machine learning system ICME Labs uses for the language-model stages of a policy check. Open source.

Read on arXiv Code on GitHub
PQCrypto 2026 Post-quantum commitments
Advances: proof stays verifiable as cryptography ages

ABBA: Lattice-based Commitments from Commutators

Alberto Centelles, Andrew Mendelsohn

A receipt is only as good as the cryptography it will be checked against years later. ABBA is a linearly homomorphic commitment scheme built from lattices, designed as a drop-in replacement for the widely used Ajtai commitment. Lattice-based schemes are the accepted path to post-quantum security, which is what lets a receipt issued today remain verifiable after current elliptic-curve assumptions fall. Published at PQCrypto 2026, the peer-reviewed conference on post-quantum cryptography.

Read on IACR ePrint
IACR ePrint, 2026 Post-quantum zero knowledge
Advances: rules and data stay hidden, under post-quantum assumptions

LatticeBlindFold: A Lattice-Based Analogue of NovaBlindFold

Luca Dall'Ava (ICME Labs)

Folding schemes are what make proofs of long computations affordable, but nearly every lattice-based folding scheme leaks information about the data being proved. That leak is fatal in regulated work, where the policy and the client file are confidential. LatticeBlindFold adds a blinding step to lattice-based folding so the proof reveals nothing about its inputs, under post-quantum assumptions. It is the lattice counterpart of the NovaBlindFold technique ICME Labs uses today to hide model weights and policy data in agent action receipts.

Read on IACR ePrint
Talks

Where this work has been presented.

September 2, 2026. Sunnyvale, California.

Google Web3 AI Summit 2026

Hosted by Google Cloud and Google DeepMind. ICME co-founder Wyatt Benno joined the Evolution of ZK panel with speakers from Coinbase, a16z crypto, and Stanford, moderated by Google Cloud.

Event page
2026. Federal Reserve Bank of New York.

New York Fed Innovation Conference

ICME co-founder Houman Shadab joined the panel on turning law into verified code, alongside leading scholars in computational law.

Watch at newyorkfed.org
Built on

Peer-reviewed cryptography, not proprietary claims.

The proof systems underneath ICME Labs' agent action receipts are published and independently reviewed: Jolt and Lasso (EUROCRYPT 2024), Nova and HyperNova (CRYPTO 2022 and 2024), Spartan (CRYPTO 2020), and Automated Reasoning checks (CAV 2026). Our own contributions extend them; they do not replace them.

Go deeper

Read the docs, or ask us directly.

For engineers

See how the research ships

The docs cover policy compilation, the solver, and receipt verification as they run in production today.

Read the docs
For research and partnership questions

Talk to the authors

Preprints, collaboration, and questions about what the proofs cover and what they do not.

Book a call