# AI verification proposal

A proposal built with the Proposal Explorer of the AI Verification Tech Map (https://trustbutveri.fyi/), from its records of 2026-10-09. Interactive version: https://trustbutveri.fyi/explorer/?mechanisms=M-0004,M-0014&cols=tested

How to read it: a claim is something one party wants to verify about another's AI hardware or software. A mechanism is a general technique for verifying claims; it is "aimed at" a claim when that is its direct purpose, and "supporting" when it contributes without being aimed at it. A claim is addressed when a mechanism in the proposal is aimed at it and is not excluded by the filters; addressed does not mean verified, so check that mechanism's development status, security evidence and findings. Definitions: https://trustbutveri.fyi/about/methodology/ (roles, properties and findings) and https://trustbutveri.fyi/about/readiness/ (development status).

## Filters

Filters apply to mechanisms only and describe the setting the proposal is for.

None set. Every mechanism on the map was available.

## Overview

One row per mechanism, read from its record. Open failures: critical / significant / minor. The last three columns are the editors' reading of what the verifier sees. Findings are grouped as known failures, scope limitations and open questions. Only known failures count as failures. Counts are an inventory of published findings, not a risk score.

| Mechanism | Development | Security evidence | Prover | Attack testing | Hardware | Open failures | Weights | Inputs and outputs | Training data |
| --- | --- | --- | --- | --- | --- | --- | --- | --- | --- |
| Zero-knowledge proofs of inference | Research demonstration | Published attack testing | Adversarial | Independent red-team | None | 0 / 0 / 0 | hidden | shown | not involved |
| Bandwidth limits and compartmentalization | Research demonstration | Published security analysis | Adversarial | Analysis | Retrofit device | 0 / 2 / 0 | not involved | not involved | not involved |

## Claims

No claims chosen.

## Mechanisms

### Zero-knowledge proofs of inference

A prover produces a cryptographic proof that an output came from running a committed model on a given input, without revealing the weights. ([Zero-knowledge proofs of inference](https://trustbutveri.fyi/mechanisms/zk-proofs-of-inference/))

- Assessment: mechanism family.
- Development: Research demonstration (legacy code R2), assessed for proving a language model's output follows from committed weights, against a cheating prover.
- Security evidence: Published attack testing. Independent evaluation: unassessed. Formal proof: unassessed. Deployment assurance: unassessed.
- Claims in this proposal: none of them.
- Threat model: adversarial prover. Hardware: none. Prover cooperation: required. Attack testing: independent red-team. Category: Cryptographic & computational.
- What the verifier sees: model weights hidden; inputs and outputs shown; training data not involved. The weights stay committed and hidden; the verifier knows each input and output it checks.

### Bandwidth limits and compartmentalization

Capping or removing network links between groups of AI chips, so each group can serve models but large training runs across groups become far slower. ([Bandwidth limits and compartmentalization](https://trustbutveri.fyi/mechanisms/bandwidth-limits-and-compartmentalization/))

- Assessment: mechanism family.
- Development: Research demonstration (legacy code R2), assessed for monitoring inter-node traffic with operator-run software on four GPUs.
- Security evidence: Published security analysis. Independent evaluation: unassessed. Formal proof: unassessed. Deployment assurance: unassessed.
- Claims in this proposal: none of them.
- Threat model: adversarial prover. Hardware: retrofit device. Prover cooperation: required. Attack testing: analysis. Category: Isolation & system architectures.
- What the verifier sees: model weights not involved; inputs and outputs not involved; training data not involved. Caps traffic between groups of chips; it does not read the traffic's content.


## Properties

**Built for an adversarial prover**

- Zero-knowledge proofs of inference
- Bandwidth limits and compartmentalization

**No new hardware needed**

- Zero-knowledge proofs of inference


## Attack testing

Attack testing records published testing for this use. It does not by itself show independent review, a formal proof or that a deployed system is secure.

**Testing history**

- Zero-knowledge proofs of inference: Independent red-team
- Bandwidth limits and compartmentalization: Analysis


## Limits

**Open significant failures**

- Operator control of pod routing collapses the bound (known failure, theoretical argument, in Bandwidth limits and compartmentalization; https://trustbutveri.fyi/mechanisms/bandwidth-limits-and-compartmentalization/evidence/flaws/2/) [14]. Lucid's analysis finds that if the operator can freely assign pods to routers, it could dedicate a whole cell of 100 or more pods to one pipeline stage. The bound then falls to about 90–220x uncompressed and as low as about 25x with compression. The proposed mitigation, auditor-controlled random assignment that is periodically re-randomized, has not been implemented.
- Parallel scale-up switches are hard enforcement points (known failure, theoretical argument, in Bandwidth limits and compartmentalization; https://trustbutveri.fyi/mechanisms/bandwidth-limits-and-compartmentalization/evidence/flaws/5/) [16]. In GB200 topologies, GPUs reach GPUs in other nodes through NVSwitches without a NIC on the path. Amodo notes that limits are hard to enforce there because many switches work in parallel, so compromising one or two would bypass the limit.

**Scope limitations**

- The proof covers a fixed-point approximation, not the floating-point model (scope limitation, open question, in Zero-knowledge proofs of inference; https://trustbutveri.fyi/mechanisms/zk-proofs-of-inference/evidence/flaws/1/) [1][5][7][11]. Current ZK inference systems prove a quantised version of the network. zkLLM scales values by 2^16 and reports small perplexity changes. Attestable reports quantising matrix multiplications to 8-bit integers while proving other operations in floating point. A verifier therefore learns about the proof-friendly variant, and must separately accept that this variant is the declared model. Trail of Bits built a ResNet-18 backdoor that is dormant in the full-precision model and active after ezkl's quantisation; whether it persists through proving was left for further investigation. A verification system design notes that ZKPs can emulate floating-point operations. Rounding makes floating-point results depend on summation order, so bit-for-bit replay of an accelerator's results needs its original reduction tree. The report calls emulating that tree inside a ZKP an open, intricate problem and asks what it would cost.
- A proof speaks only for the computations that were proven (scope limitation, theoretical argument, in Zero-knowledge proofs of inference; https://trustbutveri.fyi/mechanisms/zk-proofs-of-inference/evidence/flaws/2/) [12]. Attestable writes that "a proof of some computation is not a proof of all computation", and that a proof cannot discover a datacenter that was never declared. Proofs of inference do not by themselves show that no other workload ran on the same or other hardware.

  Related mechanism: Proofs of useful work for capacity accounting (R1, not in the proposal). The record names proof-of-work accounting as the kind of compute accounting needed to show that proven inference was the only work done.
- The model architecture is disclosed (scope limitation, theoretical argument, in Zero-knowledge proofs of inference; https://trustbutveri.fyi/mechanisms/zk-proofs-of-inference/evidence/flaws/3/) [1][3]. ZKML "requires that the model architecture (but not weights) is revealed", and zkLLM assumes a publicly known model structure. Architecture can be commercially sensitive.
- Proofs do not bind computational effort (Hollow-LLM) (scope limitation, demonstrated attack, in Zero-knowledge proofs of inference; https://trustbutveri.fyi/mechanisms/zk-proofs-of-inference/evidence/flaws/4/) [10]. Researchers at the University of Southern California show that a proof of inference certifies that an output is consistent with committed weights under the declared architecture, but not how much computation produced it. In their Hollow-LLM attack, a provider keeps the declared architecture and parameter count but commits to "ghost weights". Some layers pass their inputs through unchanged, and wide layers carry the signal in a small subspace, so a much smaller inner model does the real work. The ghost weights satisfy the verification circuit and yield valid proofs.

  The authors ran the attack with the proof procedure of zkGPT, a separate ZK inference system, on a 6-layer, 512-dimensional transformer declared as up to 12 layers and 1,024 dimensions. Outputs were identical to the inner model's, and serving cost stayed at the inner model's level. An honest model of the declared size cost 2.4 times as much to prefill and 3.1 times as much to decode. Proving cost still grew with the declared architecture.

  The authors note that results may be served before any proof, with the provider building the witness only when a call is selected for audit. They describe their constructions as "compatible with state-of-the-art zkLLM pipelines", and state that the attack does not imply a flaw in the proof system itself. They propose challenge-based audits and ablation tests, which raise the cost of cheating but give no guarantee.
- Undeclared local storage raises per-pod capacity (scope limitation, theoretical argument, in Bandwidth limits and compartmentalization; https://trustbutveri.fyi/mechanisms/bandwidth-limits-and-compartmentalization/evidence/flaws/3/) [14]. More memory or storage per pod helps an adversary. Lucid requires per-pod storage to be declared, capped and physically inspected.
- Training within one pod is not covered (scope limitation, open question, in Bandwidth limits and compartmentalization; https://trustbutveri.fyi/mechanisms/bandwidth-limits-and-compartmentalization/evidence/flaws/4/) [14]. Lucid's bounds concern pre-training models larger than the pods are sized for. Training models that fit in one pod, fine-tuning and reinforcement-learning post-training within one pod are outside the modelled threat.

**Open questions**

- Low-communication training reduces the bandwidth training needs (open question, theoretical argument, in Bandwidth limits and compartmentalization; https://trustbutveri.fyi/mechanisms/bandwidth-limits-and-compartmentalization/evidence/flaws/1/) [14][17][18]. DiLoCo matched fully synchronous training on 8 workers while communicating 500 times less. Rahman writes that this family of methods theoretically allows large-scale training with less than 100 Mbps. Lucid includes these methods in its bounds, but notes that extreme activation compression, architectures with unusually small inter-layer widths, or modular paradigms could erode the margin.


## Possible additions

Mechanisms on the map, not in the proposal, that the records connect to an unaddressed or partly addressed claim, an open failure or a dependency. Pointers, not recommendations: each brings its own readiness level and findings, and none is claimed to close a failure.

- **Tamper evidence for verifier devices** (Research demonstration (legacy code R2), assessed for detecting probing of proposed verifier hardware, using server and electronics prototypes as evidence)
  - Bandwidth limits and compartmentalization waits on it: Shaping devices and routing assignments must be trusted by both parties; Amodo has not yet fully analysed resilience to a compromised DPU.
- **Network taps and certifiers** (Proposed (legacy code R1), assessed for committing a complete record of cluster traffic, so declared inference can be checked)
  - Bandwidth limits and compartmentalization waits on it: The verifier must know that all traffic leaving a pod crosses the capped, monitored links.
- **Proofs of useful work for capacity accounting** (Proposed (legacy code R1), assessed for bounding the spare capacity of declared hardware that could run training)
  - Zero-knowledge proofs of inference waits on it: Showing that proven inference was the only work done needs a compute-accounting mechanism such as proof-of-work accounting, which is only proposed.


## Dependencies

**Missing prerequisites**

- Tamper evidence for verifier devices (Research demonstration (legacy code R2), assessed for detecting probing of proposed verifier hardware, using server and electronics prototypes as evidence), needed by Bandwidth limits and compartmentalization

**Blockers**

- Zero-knowledge proofs of inference: Proving takes about 13 minutes (803 seconds) per 2,048-token forward pass of a 13B model on one A100, and a verification system design calls the overhead heavy. (performance & compatibility) [1][11]
- Zero-knowledge proofs of inference: ZKML and zkLLM prove fixed-point arithmetic, and a verification system design calls emulating an accelerator's original floating-point reduction tree inside a zero-knowledge proof, which bit-for-bit replay needs, an open and intricate problem whose cost is also unsettled. (performance & compatibility) [1][3][11]
- Zero-knowledge proofs of inference: zkLLM's code is unaudited, interactive and archived; the one audited ZK inference library, ezkl, had high-severity circuit soundness bugs before its fixes. (adversarial validation) [2][7]
- Zero-knowledge proofs of inference: Showing that proven inference was the only work done needs a compute-accounting mechanism such as proof-of-work accounting, which is only proposed. (coverage & hidden compute; waits on Proofs of useful work for capacity accounting) [12]
- Bandwidth limits and compartmentalization: No cap that a verifier can check has been implemented or red-teamed. (adversarial validation) [14]
- Bandwidth limits and compartmentalization: The verifier must know that all traffic leaving a pod crosses the capped, monitored links. (coverage & hidden compute; waits on Network taps and certifiers) [11]
- Bandwidth limits and compartmentalization: Shaping devices and routing assignments must be trusted by both parties; Amodo has not yet fully analysed resilience to a compromised DPU. (hardware trust; waits on Tamper evidence for verifier devices) [14][16]
- Bandwidth limits and compartmentalization: Advances in low-communication training could shrink the margin that the cap enforces. (capacity bounds) [14][17][18]


## What the verifier sees

- Model weights: shown by none; depends on the design for none; hidden by Zero-knowledge proofs of inference; not involved in Bandwidth limits and compartmentalization; unspecified for none.
- Inputs and outputs: shown by Zero-knowledge proofs of inference; depends on the design for none; hidden by none; not involved in Bandwidth limits and compartmentalization; unspecified for none.
- Training data: shown by none; depends on the design for none; hidden by none; not involved in Zero-knowledge proofs of inference and Bandwidth limits and compartmentalization; unspecified for none.

## Implementations

- Zero-knowledge proofs of inference: [Attestable zero-knowledge inference prover](https://trustbutveri.fyi/implementations/attestable-zk-inference/) (R1, product); [EZKL](https://trustbutveri.fyi/implementations/ezkl/) (R2, product); [Low-trust AI compute verification system overview](https://trustbutveri.fyi/implementations/low-trust-compute-verification-system-overview/) (R1, proposed architecture); [zkLLM](https://trustbutveri.fyi/implementations/zkllm/) (R2, research prototype)
- Bandwidth limits and compartmentalization: [AI 2040 inference-only verification stack](https://trustbutveri.fyi/implementations/ai-2040-inference-only-verification-plan/) (R1, proposed architecture); [RAND secure inference data center (SIDC) design](https://trustbutveri.fyi/implementations/rand-secure-inference-data-centers/) (R1, proposed architecture)

## Sources

1. zkLLM: Zero Knowledge Proofs for Large Language Models, H. Sun et al. (2024). https://doi.org/10.1145/3658644.3670334
2. zkllm-ccs2024: code for zkLLM: Zero Knowledge Proofs for Large Language Models, H. Sun (2024). https://github.com/jvhs0706/zkllm-ccs2024
3. ZKML: An Optimizing System for ML Inference in Zero-Knowledge Proofs, B.-J. Chen et al. (2024). https://doi.org/10.1145/3627703.3650088
4. NanoZK: Privacy-Preserving Verifiable Inference for Large Language Models via Layerwise Zero-Knowledge Proofs, Z. Wang (2026). https://arxiv.org/abs/2603.18046
5. Proving LLMs at Scale, Attestable (2026). https://attestable.com/blog/proving-llms-scale
6. Verifiable evaluations of machine learning models using zkSNARKs, T. South et al. (2024). https://arxiv.org/abs/2402.02675
7. Zkonduit EZKL Security Assessment, F. Casal et al. (2025). https://github.com/trailofbits/publications/blob/master/reviews/2025-03-zkonduit-ezkl-securityreview.pdf
8. DeepProve-1: The First zkML System to Prove a Full LLM Inference, Lagrange Labs (2025). https://lagrange.dev/blog/deepprove-1
9. Lagrange-Labs/deep-prove (GitHub repository), Lagrange Labs (2026). https://github.com/Lagrange-Labs/deep-prove
10. Hollow-LLM Attack: Computationally Trivial Weights in Zero-Knowledge Verification of LLM Inference, C. Gong et al. (2026). https://arxiv.org/abs/2607.28884
11. A System Overview for Near-Term, Low-Trust AI Compute Verification, N. Cankaya (2026). https://intelligence.org/wp-content/uploads/2026/06/A-system-overview-for-near-term-low-trust-AI-compute-verification.pdf
12. Pacing AI Requires Proof, Attestable (2026). https://attestable.com/blog/pacing-ai-requires-proof
13. Verification Plan, R. Dean (2026). https://ai-2040.com/supplements/verification-plan
14. Traffic Shaping for Workload Classification, Lucid Computing (2026). https://lucidcomputing.substack.com/p/traffic-shaping-for-workload-classification
15. De-risking Interconnect Limits for AI Verification, A. Scher et al. (2026). https://techgov.intelligence.org/blog/de-risking-interconnect-limits-for-ai-verification
16. The Tray as a Bandwidth Boundary, Amodo Design (2026). https://amododesign.com/notes/2026-03-16-dpu-bandwidth-limiter/
17. DiLoCo: Distributed Low-Communication Training of Language Models, A. Douillard et al. (2024). https://arxiv.org/abs/2311.08105
18. Does Distributed Training Undermine Compute Governance?, R. Rahman (2026). https://arxiv.org/abs/2605.29359