Proposal Explorer

Add claims to see which ones the mechanisms address.

In production

• R3 Model identity attestation for showing clients that the served weights match a committed hash
• R3 Sampled inference recomputation for checking untrusted workers' activations against the declared model, prompt and precision

Built for an adversarial prover

Sampled inference recomputation

No new hardware needed

Model identity attestation and Sampled inference recomputation

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

• Model identity attestation, Tinfoil model identity (Modelwrap): Analysis
• Sampled inference recomputation: Analysis

Open critical flaws

• Inherits attacks on the underlying TEEs in Tinfoil model identity (Modelwrap)

  Critical when model identity must hold against an operator with physical access to an affected host. Tinfoil documents physical attacks as an enclave limitation. These are hardware-class demonstrations, not a published break of Modelwrap or Tinfoil's deployed verification chain.

  Tinfoil's model commitment and boot-time GPU check depend on the CPU attestation. The TEE findings distinguish Intel TDX forgery on DDR5, AMD SEV-SNP forgery on DDR4 in Battering RAM, and software-only RMPocalypse on platforms lacking AMD's fixes. TEE.fail recovered a guest OpenSSL key on AMD, not an AMD attestation key. Its GPU relay demonstration used an H100 with forged TDX evidence; it does not establish the same result for Tinfoil's H200 or B200 configurations. ^{2 3 6 7 8 9}

  Response: Tinfoil acknowledges the physical-access boundary. The TEE.fail authors report that Intel and AMD treat interposer attacks as outside their threat models and recommend physically secure servers. AMD reports firmware fixes for RMPocalypse.

  Demonstrated attack · Critical · Open · Inherited finding. On the record · Related finding in TEE remote attestation for AI workloads

Family findings

• Model identity attestation

  Context for Tinfoil model identity (Modelwrap). These findings concern the family or other implementations; applicability must be checked against their stated scope.

  • Underlying attestation can be forged or relayed in Model identity attestation

    Critical for the enclave route against an operator with physical access to affected hardware, or control of an unpatched SEV-SNP hypervisor. It does not apply to the recomputation route. PAL*M excludes physical attacks, and Tinfoil acknowledges this boundary.

    The enclave route inherits the platform-specific TEE attestation failures. Intel TDX forgery and H100 relay were demonstrated with physical access and host control. Battering RAM defeated AMD SEV-SNP attestation on DDR4 servers; RMPocalypse did so from malicious host software on platforms without AMD's fixes. These demonstrate failures of the trust roots, not of each model-commitment protocol. ^{2 6 7 8 9 10}

    Response: The TEE.fail authors report that physical interposer attacks are outside Intel's and AMD's threat models. AMD reports fixes for RMPocalypse.

    Demonstrated attack · Critical · Open · Inherited finding. On the record · Related finding in TEE remote attestation for AI workloads

    Related mechanism R1 Hardware-enabled guarantees (flexHEG) and guarantee processors: A tamper-protected enclosure around the chip is the proposed answer when the party that holds the hardware may attack it physically. A pointer, not evidence that this flaw is mitigated. Add

  • Launch-state attestation does not by itself cover weights loaded later in Model identity attestation

    Attestation measures launch state, and weights are read from disk after boot. A signature checked at load time does not stop a malicious hypervisor from altering the disk afterwards. Tinfoil reports mitigating this with dm-verity checks on every read. Unmeasured runtime configuration remains a general risk. ^{1 11}

    Theoretical argument · Significant · Mitigated. On the record

  • For private models, a user can confirm consistency but not content in Model identity attestation

    When weights are not published, users can check that the same root hash is served each time, but not what the model is. Pairing the hash with an attested evaluation, as in Attestable Audits, is one proposed remedy. ^{1 12}

    Open question · Significant · Open. On the record

  • Recomputation depends on trusted logging and randomness, and its tolerance leaves a covert channel in Model identity attestation

    The recomputation variant assumes that every input, output and seed is logged correctly, and that the attacker can neither predict nor manipulate which messages are sampled for verification. Legitimate nondeterminism concentrates at a few token positions, and slow leaks within the tolerated slack remain possible. An independent study showed that an adversary who controls the prompts roughly doubles the bits leaked per token, reducing the exfiltration slowdown from 146–254 times under benign prompts to 60–118 times. The attack targets the exfiltration bound, not the check that outputs match the declared model. ^{13 14}

    Demonstrated attack · Significant · Open. On the record

    Related mechanism R1 Network taps and certifiers: Taps are proposed to copy and hash traffic on the monitored links, reducing reliance on the prover's own log. This still depends on the monitored boundary and trusted capture. A pointer, not evidence that this flaw is mitigated. Add

    Related mechanism R3 Deterministic and bit-exact inference: Bit-exact inference would remove the numerical tolerance that leaves this channel. A pointer, not evidence that this flaw is mitigated. Add

Open significant flaws

5 flaws in 2 mechanisms

• Side channels, I/O leakage and denial of service are outside enclave protection in Tinfoil model identity (Modelwrap)

  Tinfoil's documentation lists timing, power and electromagnetic side channels, host observation of access patterns and I/O, denial of service, supply-chain compromise and rollback as limitations. ²

  Theoretical argument · Significant · Open. On the record

• Private models can be checked only for consistency in Tinfoil model identity (Modelwrap)

  For unpublished weights, the root hash appears in the attestation without the weights being exposed. Users can then confirm only that they get the same model each time. ¹

  Open question · Significant · Open. On the record

• Tolerance for numerical noise leaves a covert channel in Sampled inference recomputation

  Schemes that accept approximate matches can put an upper bound on an adversary's covert bandwidth, but they cannot close the channel. The weight-exfiltration detector cut exfiltratable information to under 0.5%, not to zero, on a 30-billion-parameter mixture-of-experts model under benign prompt traffic. Its authors called the channel's size under adversarial prompts an open empirical question. An independent study showed that an adversary who controls the prompts roughly doubles the bits leaked per token. Across six models, that cut the slowdown from 146–254 times under benign prompts to 60–118 times. The attack widens the exfiltration bound. It does not target the check that outputs match the declared model. ^{13 14 24}

  Demonstrated attack · Significant · Open. On the record

  Related mechanism R3 Deterministic and bit-exact inference: Bit-exact inference would remove the numerical tolerance if exact replay can be deployed with the required weights and configuration. A pointer, not evidence that this flaw is mitigated. Add

• Only recorded traffic is checked in Sampled inference recomputation

  Recomputation checks that recorded, declared workloads are correct. It cannot show that the record is complete. The published schemes do not cover hidden workloads run on the same compute, or substituted work. Rinberg et al. say their exfiltration-detection scheme cannot stand alone. ^{13 25}

  Theoretical argument · Significant · Open. On the record

  Related mechanism R1 Network taps and certifiers: Taps copy and hash all traffic on the monitored links, which bears on whether the traffic record is complete. They do not show what else ran on the same chips. A pointer, not evidence that this flaw is mitigated. Add

• Some inference optimizations are not covered in Sampled inference recomputation

  TOPLOC's authors state that it cannot detect speculative decoding in which a cheaper model does the decoding. They did not test whether it distinguishes types of key-value (KV) cache compression. DiFR was evaluated only on sampling from a single model. Its authors sketch an extension to one speculative-decoding algorithm but do not test it. ^{15 16}

  Theoretical argument · Significant · Open. On the record

Open minor flaws

1 mechanism with minor findings

• Mixed hardware widens the honest baseline in Sampled inference recomputation

  When honest reference runs span different GPU types, the spread of benign scores grows. In DiFR's tests on Qwen3-30B-A3B, pooling A100 and H200 runs left Token-DiFR unable to separate the two smallest tested changes, a temperature of 1.1 instead of 1.0 and a simulated top-2 sampling bug, at the target false-positive rate, while cross-entropy separated them. Matched provider and verifier environments, or pooling that weights rare large deviations, restored detection. ¹⁵

  Open question · Minor · Open. On the record

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

• R3 Deterministic and bit-exact inference Add
  • Bears on the open significant flaw “Tolerance for numerical noise leaves a covert channel” in Sampled inference recomputation. Bit-exact inference would remove the numerical tolerance if exact replay can be deployed with the required weights and configuration.
• R1 Network taps and certifiers Add
  • Bears on the open significant flaw “Only recorded traffic is checked” in Sampled inference recomputation. Taps copy and hash all traffic on the monitored links, which bears on whether the traffic record is complete. They do not show what else ran on the same chips.
  • Sampled inference recomputation waits on it. In tap-based retrofit designs, recording all inference traffic needs network taps and recomputation servers that can ingest it, in the worst case one recomputation-server network interface per inference front-end interface.
• R3 TEE remote attestation for AI workloads Add
  • Model identity attestation waits on it. The underlying TEE attestation does not resist attackers with physical access to the host.

Missing prerequisites

• R3 TEE remote attestation for AI workloads needed by Model identity attestation Add

Blockers

7 blockers recorded

• Model identity attestation
  • The underlying TEE attestation does not resist attackers with physical access to the host. Hardware trust. Waits on TEE remote attestation for AI workloads ^{2 6}
  • No independent evaluation of the model-identity chain has been published. Adversarial validation
• Sampled inference recomputation
  • In tap-based retrofit designs, recording all inference traffic needs network taps and recomputation servers that can ingest it, in the worst case one recomputation-server network interface per inference front-end interface. Coverage & hidden compute. Waits on Network taps and certifiers ^{21 27}
  • In retrofit designs, the recomputation server must sit inside the prover's data centre, possibly under the prover's physical control, and still be protected from a compromised provider, which Amodo rates 'not on track'. Hardware trust ^{13 21}
  • No independent red-team of a recomputation consistency check has been published (the one independent attack study targets the weight-exfiltration bound), and Amodo rates recomputation red-teaming 'not started'. Adversarial validation ^{14 21}
  • Tolerance-based checks need calibration on trusted hardware and exact knowledge of the provider's sampling procedure, and in one prototype a sampling-implementation mismatch produced large spurious differences. Performance & compatibility ^{15 20}
  • The verifier needs the model weights, so checking a closed-weights model requires a trusted, confidential recomputation environment, which the retrofit designs place inside the prover's facility. Privacy & leakage ^{15 25 26}

From the family or selected implementation's record. ?

Model weights

Depends on the design for Model identity attestation and Sampled inference recomputation.

Inputs and outputs

Depends on the design for Sampled inference recomputation.

Unspecified for Model identity attestation. Check the implementation record.

Training data

Not involved: Model identity attestation and Sampled inference recomputation.

Model identity attestation

• R2 Attestable Audits Research prototype, University of Cambridge
• R2 PAL*M Research prototype, University of Waterloo
• R3 Tinfoil model identity (Modelwrap) Product, Tinfoil

Sampled inference recomputation

• R1 AI 2040 inference-only verification stack Proposed architecture, AI Futures Project
• R2 DiFR (Divergence From Reference) Research prototype
• 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)
• R3 TOPLOC Open-source project, Prime Intellect

1. How Tinfoil Proves Exactly What Model Is Running, Tinfoil Team (2026). Original
2. A primer on secure enclaves, Tinfoil (2026). Original
3. Backend infrastructure, Tinfoil (2026). Original
4. How verification works in Tinfoil, Tinfoil (2026). Original
5. modelwrap: Reproducible dm-verity read-only image of Huggingface models, Tinfoil (2026). Original
6. TEE.fail: Breaking Trusted Execution Environments via DDR5 Memory Bus Interposition, J. Chuang et al. (2026). Original
7. Battering RAM: Low-Cost Interposer Attacks on Confidential Computing via Dynamic Memory Aliasing, J. De Meulemeester et al. (2026). Original
8. RMPocalypse: How a Catch-22 Breaks AMD SEV-SNP, B. Schlüter & S. Shinde (2025). Original
9. SEV-SNP RMP Initialization Vulnerability (AMD-SB-3020), AMD (2025). Original
10. PAL*M: Property Attestation for Large Generative Models, P. Chantasantitam et al. (2026). Original
11. On TEEs for Privacy-Preserving Monitoring in AI Governance, Gloria Z (2026). Original
12. Attestable Audits: Verifiable AI Safety Benchmarks Using Trusted Execution Environments, C. Schnabl et al. (2025). Original
13. Verifying LLM Inference to Detect Model Weight Exfiltration, R. Rinberg et al. (2025). Original
14. Adversarial Entropy Inflation Against Gumbel-Based Inference Verification, N. Kezins (2026). Original
15. DiFR: Inference Verification Despite Nondeterminism, A. Karvonen et al. (2025). Original
16. TOPLOC: A Locality Sensitive Hashing Scheme for Trustless Verifiable Inference, J. M. Ong et al. (2025). Original
17. PrimeIntellect-ai/toploc (GitHub repository), Prime Intellect (2025). Original
18. INTELLECT-2: A Reasoning Model Trained Through Globally Decentralized Reinforcement Learning, Prime Intellect Team et al. (2025). Original
19. adamkarvonen/difr (GitHub repository), A. Karvonen (2025). Original
20. Scaling Recomputation Inference Verification, Amodo Design (2026). Original
21. AI 2040 Plan A — Verification SITREP, Amodo Design (2026). Original
22. SYNTHETIC-2 Release: Four Million Collaboratively Generated Reasoning Traces, Prime Intellect (2025). Original
23. An Inference Verification Prototype — Stage 1, Amodo Design (2026). Original
24. Bit-Exact AI Inference Verification Without Performance Tradeoffs, N. Cankaya (2026). Original
25. Example Schemes for Verifying High-Stakes AI Agreements, Amodo Design (2026). Original
26. A System Overview for Near-Term, Low-Trust AI Compute Verification, N. Cankaya (2026). Original
27. Verification Plan, R. Dean (2026). Original