Proposal Explorer

Add claims to see which ones the mechanisms address.

Built for an adversarial prover

Zero-knowledge proofs of inference and Network taps and certifiers

No new hardware needed

Zero-knowledge proofs of inference

Flaws since mitigated

• Circuit and contract bugs allowed forged proofs in Zero-knowledge proofs of inference ³

Published attempts to break a system, including those that found failures. Testing history does not show that open flaws are resolved.

• Zero-knowledge proofs of inference, EZKL: Independent red-team

Family findings

• Zero-knowledge proofs of inference

  Context for EZKL. These findings concern the family or other implementations; applicability must be checked against their stated scope.

  • The proof covers a fixed-point approximation, not the floating-point model in Zero-knowledge proofs of inference

    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 calls floating-point emulation in ZKPs an open problem. ^{3 4 5 6}

    Open question · Significant · Open. On the record

  • A proof speaks only for the computations that were proven in Zero-knowledge proofs of inference

    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. ⁷

    Theoretical argument · Significant · Open. On the record

    Related mechanism R1 Proofs of useful work for capacity accounting: The record names proof-of-work accounting as the kind of compute accounting needed to show that proven inference was the only work done. A pointer, not evidence that this flaw is mitigated. Add

  • The model architecture is disclosed in Zero-knowledge proofs of inference

    ZKML "requires that the model architecture (but not weights) is revealed", and zkLLM assumes a publicly known model structure. Architecture can be commercially sensitive. ^{4 8}

    Theoretical argument · Minor · Open. On the record

  • Proofs do not bind computational effort (Hollow-LLM) in Zero-knowledge proofs of inference

    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. ⁹

    Demonstrated attack · Significant · Open. On the record

• Network taps and certifiers

  Context for AI 2040 inference-only verification stack. These findings concern the family or other implementations; applicability must be checked against their stated scope.

  • Output nondeterminism leaves covert capacity in Network taps and certifiers

    Hashing cannot remove information hidden in the outputs themselves. The Secure Gateway Device paper estimates that about 0.1 bit per token remains even with seed-synchronized replay checks. For a 200k-GPU inference cluster at full load (2,000 tokens per GPU per second), that is about 40 Mbit/s of covert egress, enough to move a 1 TB model in under three days. The paper names this the core remaining challenge and points to deterministic replay or active scrubbing of hardware-induced entropy. An independent study found that an adversary who chooses the prompts roughly doubles the bits leaked per token under Gumbel-based inference verification; see Bounding unexplained information in outputs. ^{16 17}

    Theoretical argument · Significant · Open. On the record

    Related mechanism R3 Deterministic and bit-exact inference: Deterministic replay is one of the two remedies the flaw's source names. A pointer, not evidence that this flaw is mitigated. Add

    Related mechanism R2 Bounding unexplained information in outputs: Bounds the hidden information outputs can carry by measuring what the declared computation fails to predict. A pointer, not evidence that this flaw is mitigated. Add

  • Some links cannot be passively tapped in Network taps and certifiers

    Cankaya notes that copper-connected scale-up domains (for example NVL72 racks and TPU v7 cubes) are much harder to tap than fibre, and that optical budgets make passive taps impractical on 400GBASE-SR8 multimode links. Amodo found no taps advertised for 53 GBaud links as of May 2026. ^{18 19}

    Open question · Significant · Open. On the record

  • Encrypted fabrics hide plaintext from both parties in Network taps and certifiers

    Cankaya notes that with TEE-protected sessions whose keys are ephemeral and managed inside the TEE, neither the operator nor the manufacturer can recover session keys after the session, so tapped traffic could not be opened for recomputation. For other encrypted fabrics, the operator can retain keys. ¹⁸

    Open question · Significant · Open. On the record

  • Residual side channels in simple passive setups in Network taps and certifiers

    Amodo's analysis of its own tapped prototype lists unvalidated header fields, timing of permitted traffic and variation in response formatting as residual channels, and concludes that the passive tap must be replaced by an active one. ¹⁴

    Theoretical argument · Significant · Open. On the record

  • Completeness rests on physical monitoring left out of scope in Network taps and certifiers

    The Secure Gateway Device paper assumes the facility is physically monitored, and states that the whole architecture depends on the device being the only communication channel. It names radio emanation, power-line signalling and thermal channels as covert channels beyond that scope. ¹⁷

    Open question · Significant · Open. On the record

    Related mechanism R1 Side-channel suppression for isolated facilities: Addresses the radio, power-line and thermal channels that network-level designs leave out. A pointer, not evidence that this flaw is mitigated. Add

  • Verifier dictionary attacks on hashes in Network taps and certifiers

    Hashes of very short outputs could be brute-forced by the verifier. The paper recommends hashing at least 5 tokens together, or at least 10 if the attacker filters for likely tokens. ¹⁷

    Theoretical argument · Minor · Mitigated. On the record

Open significant flaws

4 flaws in 2 mechanisms

• Quantization can activate a backdoor dormant in the full-precision model in EZKL

  EZKL quantizes values to represent them in a finite field. Trail of Bits built a ResNet-18 whose backdoor is dormant at full precision and active after EZKL's quantization. Larger models and smaller quantization scales make the attack easier. Whether the backdoor persists through the witness and proof stages was left for further investigation. The fix was documentation of the risk. ³

  Demonstrated attack · Significant · Open. On the record

• The recomputation server must be trusted in AI 2040 inference-only verification stack

  The plan calls the integrity of the recomputation server an extremely important aspect, and its argument that sampling verifies all outputs assumes that the server's computations and outputs can be trusted. The companion page notes that the server sits inside the prover's facility, possibly under the prover's physical control, and that hardening it against integrity attacks needs significant research. Amodo rates recomputation-server security as not on track. ^{10 11 13}

  Theoretical argument · Significant · Open. On the record

• Spare compute is not verified in AI 2040 inference-only verification stack

  The plan states that it does not verify that spare compute is unused for unapproved workloads, because this seems very challenging. It relies instead on side-channel bounds and memory wipes, so that the only results that persist are verified inference outputs. ¹⁰

  Theoretical argument · Significant · Open. On the record

• A recomputation family degrades against prompt-controlling adversaries in AI 2040 inference-only verification stack

  The plan's companion page names DiFR among the recomputation schemes being tested. An independent study found that Gumbel-based inference verification, the family that includes Token-DiFR, leaks roughly twice as many bits per token when the adversary chooses prompts that disrupt the text's structure, across six models of 1 to 32 billion parameters. The slowdown it imposes on a hidden-signalling adversary falls from 146–254 times under benign prompts to 60–118 times. The attack weakens the bound on hidden information in outputs, which the plan relies on to keep undeclared results from leaving. ^{10 11 15 16}

  Demonstrated attack · Significant · Open. On the record

Not yet demonstrated

R1 Network taps and certifiers

Mechanisms on the map that are not in the proposal. Pointers, not recommendations. ?

• R2 Tamper evidence for verifier devices Add
  • Network taps and certifiers waits on it. Checking that taps are correctly installed and stay in place at scale is not a solved problem, and hardening the recomputation server inside the prover's facility needs significant research.
• R1 Memory wiping and proofs of secure erasure Add
  • Network taps and certifiers waits on it. Memory wiping may use existing algorithms, but hardware testing is at an early stage.
• R1 Side-channel suppression for isolated facilities Add
  • Network taps and certifiers waits on it. There is no plan yet for quickly scaling side-channel defences on a frontier cluster; only early theoretical pieces exist.
• R1 Whole-workload recomputation (reproducible packets) Add
  • Network taps and certifiers waits on it. A fully reproducible inference stack needs substantial software and tooling, and per-packet network reproducibility may need considerable software, firmware and possibly hardware work.
• R3 Sampled inference recomputation Add
  • Network taps and certifiers depends on it.

Missing prerequisites

• R3 Sampled inference recomputation needed by Network taps and certifiers Add
• R1 Whole-workload recomputation (reproducible packets) needed by Network taps and certifiers Add

Blockers

7 blockers recorded

• Zero-knowledge proofs of inference
  • Proving cost grows steeply with model size: a 250,000-parameter nanoGPT took 2,781 s to prove and needed a 219 GB proving key, which South et al. name as the main limit on model size. Performance & compatibility ²
• Network taps and certifiers
  • A fully reproducible inference stack needs substantial software and tooling, and per-packet network reproducibility may need considerable software, firmware and possibly hardware work. Performance & compatibility. Waits on Whole-workload recomputation (reproducible packets) ¹¹
  • Passive optical taps work at 400G, but the 800G and 1600G line rates now arriving in data centres are undemonstrated. Performance & compatibility. Waits on Network taps and certifiers ¹¹
  • Checking that taps are correctly installed and stay in place at scale is not a solved problem, and hardening the recomputation server inside the prover's facility needs significant research. Hardware trust. Waits on Tamper evidence for verifier devices ^{11 13}
  • There is no plan yet for quickly scaling side-channel defences on a frontier cluster; only early theoretical pieces exist. Coverage & hidden compute. Waits on Side-channel suppression for isolated facilities ¹¹
  • Memory wiping may use existing algorithms, but hardware testing is at an early stage. Coverage & hidden compute. Waits on Memory wiping and proofs of secure erasure ¹¹
  • Robust red-teaming of recomputation schemes has not started, and most algorithm development remains academic. Adversarial validation ^{11 13}

From the family or selected implementation's record. ?

Model weights

Unspecified for Zero-knowledge proofs of inference and Network taps and certifiers. Check the implementation record.

Inputs and outputs

Unspecified for Zero-knowledge proofs of inference and Network taps and certifiers. Check the implementation record.

Training data

Unspecified for Zero-knowledge proofs of inference and Network taps and certifiers. Check the implementation record.

Zero-knowledge proofs of inference

• R1 Attestable zero-knowledge inference prover Product, Attestable
• R2 EZKL Product, Zkonduit
• R1 Low-trust AI compute verification system overview Proposed architecture, Machine Intelligence Research Institute
• R2 zkLLM Research prototype, University of Waterloo

Network taps and certifiers

• R1 AI 2040 inference-only verification stack Proposed architecture, AI Futures Project
• R1 Low-trust AI compute verification system overview Proposed architecture, Machine Intelligence Research Institute
• R1 SASH confidential network logger Research prototype, Singapore AI Safety Hub (SASH)

1. zkonduit/ezkl (GitHub repository), Zkonduit Inc. (2026). Original
2. Verifiable evaluations of machine learning models using zkSNARKs, T. South et al. (2024). Original
3. Zkonduit EZKL Security Assessment, F. Casal et al. (2025). Original
4. zkLLM: Zero Knowledge Proofs for Large Language Models, H. Sun et al. (2024). Original
5. Proving LLMs at Scale, Attestable (2026). Original
6. A System Overview for Near-Term, Low-Trust AI Compute Verification, N. Cankaya (2026). Original
7. Pacing AI Requires Proof, Attestable (2026). Original
8. ZKML: An Optimizing System for ML Inference in Zero-Knowledge Proofs, B.-J. Chen et al. (2024). Original
9. Hollow-LLM Attack: Computationally Trivial Weights in Zero-Knowledge Verification of LLM Inference, C. Gong et al. (2026). Original
10. Verification Plan, R. Dean (2026). Original
11. Get Involved in Verification, AI Futures Project (2026). Original
12. Verifying international AI deals: Plan A, the state-of-play, and what you can do to help, T. Milton et al. (2026). Original
13. AI 2040 Plan A — Verification SITREP, Amodo Design (2026). Original
14. Fitting a Network TAP to our Inference Verification Prototype, Amodo Design (2026). Original
15. Verifying LLM Inference to Detect Model Weight Exfiltration, R. Rinberg et al. (2025). Original
16. Adversarial Entropy Inflation Against Gumbel-Based Inference Verification, N. Kezins (2026). Original
17. Fingerprinting All AI Cluster I/O Without Mutually Trusted Processors, N. Cankaya et al. (2026). Original
18. The Fundamentals and Feasibility of Secure Network Taps for Verifying AI Datacenter Use, N. Cankaya (2026). Original
19. Network Tapping for AI Verification: A Technical Assessment, Amodo Design (2026). Original