# AI verification proposal

A proposal built with the Proposal Explorer of the AI Verification Tech Map (https://trustbutveri.fyi/), from its records of 2026-10-08. Interactive version: https://trustbutveri.fyi/explorer/?mechanisms=M-0023,M-0014&implementations=M-0014:I-0010

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 readiness and open flaws. Readiness levels R0 to R4 describe one record's public evidence for its assessed use and are never combined. Definitions: https://trustbutveri.fyi/about/methodology/ (roles, properties and flaws) and https://trustbutveri.fyi/about/readiness/ (readiness levels).

## 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 flaws: critical / significant / minor. The last three columns are the editors' reading of what the verifier sees.

| Mechanism | Readiness | Prover | Attack testing | Hardware | Open flaws | Weights | Inputs and outputs | Training data |
| --- | --- | --- | --- | --- | --- | --- | --- | --- |
| Safeguard attestation | R2 | Semi-trusted | Analysis | Existing features | 0 / 5 / 0 | depends | depends | not involved |
| Bandwidth limits and compartmentalization / RAND secure inference data center (SIDC) design | R1 | Semi-trusted | Analysis | Retrofit device | 0 / 1 / 1 | unspecified | unspecified | unspecified |

## Claims

No claims chosen.

## Mechanisms

### Safeguard attestation

Hardware-signed evidence that an AI service ran its declared safeguards, such as a guardrail classifier or monitor, when producing a given response. ([Safeguard attestation](https://trustbutveri.fyi/mechanisms/safeguard-attestation/))

- Assessment: mechanism family.
- Readiness: R2 Demonstrated, assessed for attesting that a declared safeguard mediated a service's responses.
- Claims in this proposal: none of them.
- Threat model: semi-trusted prover. Hardware: existing features. Prover cooperation: required. Attack testing: analysis. Category: Cryptographic & computational.
- What the verifier sees: model weights depends; inputs and outputs depends; training data not involved. The enclave route signs hashes of the safeguard, request and response; a low-trust design has the verifier re-run and screen sampled requests itself.

### Bandwidth limits and compartmentalization

A RAND design for a purpose-built facility that serves already-trained AI models while protecting weights and inference data against state-level attackers. ([Bandwidth limits and compartmentalization](https://trustbutveri.fyi/mechanisms/bandwidth-limits-and-compartmentalization/))

- Assessment: selected implementation [RAND secure inference data center (SIDC) design](https://trustbutveri.fyi/implementations/rand-secure-inference-data-centers/).
- Readiness: R1 Proposed, assessed for the operator's own weight security, with no outside verification described.
- Claims in this proposal: none of them.
- Threat model: semi-trusted prover. Hardware: retrofit device. Prover cooperation: required. Attack testing: analysis. Category: Isolation & system architectures.
- What the verifier sees: model weights unspecified; inputs and outputs unspecified; training data unspecified. This Explorer has no asset-specific exposure assessment for this implementation. Check its source and deployment assumptions.


## Properties

**No new hardware needed**

- Safeguard attestation


## Attack testing

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

**Testing history**

- Safeguard attestation: Analysis
- Bandwidth limits and compartmentalization / RAND secure inference data center (SIDC) design: Analysis


## Limits

**Open significant flaws**

- Attestation shows a safeguard ran, not that it is effective (theoretical argument, in Safeguard attestation; https://trustbutveri.fyi/mechanisms/safeguard-attestation/#flaw-1) [1]. Proof of guardrail ensures that the guardrail executed, but the guardrail can still err or be jailbroken. Because the guardrail must be open source, a malicious developer can attack it with jailbreaks while still presenting a valid proof. In the authors' evaluation, Llama Guard 3 reached an F1 score of 0.56 on the unsafe class of the ToxicChat dataset. The authors state that proof of guardrail should not be interpreted or advertised as proof of safety.
- Selective attestation leaves traffic uncovered (theoretical argument, in Safeguard attestation; https://trustbutveri.fyi/mechanisms/safeguard-attestation/#flaw-2) [1][4][9]. Attestations are issued per response. In the prototype, the agent offers them when it receives high-stakes questions, so nothing shows that unattested traffic went through the same path. PAL*M's authors note that a prover could cherry-pick favourable executions, and suggest verifier-published nonces or requesting only session-level proofs. A governance analysis notes that auditors also need assurance that all activity is accounted for, since a host could start a second confidential virtual machine that bypasses monitoring.

  Related mechanism: On-chip telemetry from timing, memory and performance counters (R2, not in the proposal). On-chip counters are a proposed route to evidence about everything a chip runs, including a second virtual machine that skips the safeguard.
- Measurements may omit behaviour-relevant configuration or runtime changes (theoretical argument, in Safeguard attestation; https://trustbutveri.fyi/mechanisms/safeguard-attestation/#flaw-3) [9]. Every component that influences inference behaviour must be covered by the launch measurement, including feature flags, environment variables and invocation arguments. A launch measurement also does not show that a program keeps running as measured if the kernel is later compromised.
- Components outside the attested boundary (theoretical argument, in Safeguard attestation; https://trustbutveri.fyi/mechanisms/safeguard-attestation/#flaw-4) [1][2]. In the proof-of-guardrail experiments, the guardrail model and the agent's backend model were both reached through external APIs, and the authors leave the decision to trust those APIs to the verifier. The measured wrapper must also have no vulnerability that lets the unmeasured agent bypass the guardrail, for example by executing arbitrary commands inside the enclave. The code's README states that the enclave does not currently restrict the agent's arbitrary command execution, which could be used to bypass guardrails.
- Memory-bus interposition extracts attestation keys and forges attestations (demonstrated attack, in Safeguard attestation; https://trustbutveri.fyi/mechanisms/safeguard-attestation/#flaw-5) [1][4][7][8][9][10][11][12][13]. Inherited finding. Applies to variants using the affected Intel or AMD trust roots. PAL*M excludes physical attacks. A TDX-backed safeguard claim against a physical host attacker would be defeated, but these studies do not demonstrate a break of the AWS Nitro proof-of-guardrail prototype or of verifier-side recomputation. The TEE findings cover DDR5 attacks on Intel TDX, the H100 relay demonstration, DDR4 attacks on AMD SEV-SNP, and software-only SEV-SNP forgery before AMD's fixes. These are inherited hardware limits; a governance analysis explains why physical access matters in a treaty setting. Related finding: https://trustbutveri.fyi/mechanisms/tee-remote-attestation/#flaw-1.

  Response: Intel and AMD place the physical attack class outside their threat models, according to the researchers. AMD reports firmware fixes for RMPocalypse.

  Related mechanism: Hardware-enabled guarantees (flexHEG) and guarantee processors (R1, not in the proposal). A tamper-protected enclosure around the chip is the proposed answer when the party that holds the hardware may attack it physically.
- Everything rests on the trusted setup (theoretical argument, in RAND secure inference data center (SIDC) design; https://trustbutveri.fyi/implementations/rand-secure-inference-data-centers/#flaw-1) [14]. Reference measurements for model weights and reference data are established in a trusted setup phase. The report states that the system cannot detect compromise that happened before ingestion if the trusted setup itself is compromised.

**Family finding context**

- Context for RAND secure inference data center (SIDC) design; applicability depends on the finding's scope. Low-communication training reduces the bandwidth training needs (theoretical argument, in Bandwidth limits and compartmentalization; https://trustbutveri.fyi/mechanisms/bandwidth-limits-and-compartmentalization/#flaw-1) [15][16][17]. 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.
- Context for RAND secure inference data center (SIDC) design; applicability depends on the finding's scope. Operator control of pod routing collapses the bound (theoretical argument, in Bandwidth limits and compartmentalization; https://trustbutveri.fyi/mechanisms/bandwidth-limits-and-compartmentalization/#flaw-2) [17]. 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.
- Context for RAND secure inference data center (SIDC) design; applicability depends on the finding's scope. Undeclared local storage raises per-pod capacity (theoretical argument, in Bandwidth limits and compartmentalization; https://trustbutveri.fyi/mechanisms/bandwidth-limits-and-compartmentalization/#flaw-3) [17]. More memory or storage per pod helps an adversary. Lucid requires per-pod storage to be declared, capped and physically inspected.
- Context for RAND secure inference data center (SIDC) design; applicability depends on the finding's scope. Training within one pod is not covered (open question, in Bandwidth limits and compartmentalization; https://trustbutveri.fyi/mechanisms/bandwidth-limits-and-compartmentalization/#flaw-4) [17]. 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.
- Context for RAND secure inference data center (SIDC) design; applicability depends on the finding's scope. Parallel scale-up switches are hard enforcement points (theoretical argument, in Bandwidth limits and compartmentalization; https://trustbutveri.fyi/mechanisms/bandwidth-limits-and-compartmentalization/#flaw-5) [18]. 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.

**Open minor flaws**

- Security weakens over long operation (theoretical argument, in RAND secure inference data center (SIDC) design; https://trustbutveri.fyi/implementations/rand-secure-inference-data-centers/#flaw-2) [14]. The authors claim that the facility can withstand attacks at the OC5 level for a five-year operational period. They expect its ability to withstand long OC5 campaigns to become less robust the longer the facility remains in operation.

**Not yet demonstrated**

- Bandwidth limits and compartmentalization: R1 Proposed, assessed for the operator's own weight security, with no outside verification described


## Possible additions

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

- **On-chip telemetry from timing, memory and performance counters** (R2 Demonstrated, assessed for workload evidence from GPU counters and timing, assuming authentic measurements)
  - Bears on the open significant flaw "Selective attestation leaves traffic uncovered" in Safeguard attestation. On-chip counters are a proposed route to evidence about everything a chip runs, including a second virtual machine that skips the safeguard.
- **Hardware-enabled guarantees (flexHEG) and guarantee processors** (R1 Proposed, assessed for checking and enforcing training-compute limits on chips, against adversaries up to states)
  - Bears on the open significant flaw "Memory-bus interposition extracts attestation keys and forges attestations" in Safeguard attestation. A tamper-protected enclosure around the chip is the proposed answer when the party that holds the hardware may attack it physically.
- **Model identity attestation** (R3 In production, assessed for showing users that a service runs the declared model weights)
  - Safeguard attestation waits on it: Safeguard evidence must be bound to the model actually served, which depends on model-identity attestation.
  - Bandwidth limits and compartmentalization depends on it.
- **TEE remote attestation for AI workloads** (R3 In production, assessed for showing which software ran to a party that distrusts the operator holding the hardware)
  - Safeguard attestation waits on it: Frontier model inference typically needs several GPUs, GPU confidential computing is less mature than CPU support, and CPU inference, which an enclave prototype had to use, ran about 100 times slower than GPU inference.
  - Safeguard attestation waits on it: Trust rests on a small number of hardware vendors, and a per-CPU Intel attestation key has been extracted by physical attack.


## Dependencies

**Missing prerequisites**

- TEE remote attestation for AI workloads (R3 In production, assessed for showing which software ran to a party that distrusts the operator holding the hardware), needed by Safeguard attestation
- Model identity attestation (R3 In production, assessed for showing users that a service runs the declared model weights), needed by Safeguard attestation and Bandwidth limits and compartmentalization

**Shared foundations**

- Model identity attestation, relied on by Safeguard attestation and Bandwidth limits and compartmentalization

**Blockers**

- Safeguard attestation: No published design shows that all of a provider's traffic passes through the attested safeguard path; current evidence covers individual attested responses. (coverage & hidden compute) [1][9]
- Safeguard attestation: Frontier model inference typically needs several GPUs, GPU confidential computing is less mature than CPU support, and CPU inference, which an enclave prototype had to use, ran about 100 times slower than GPU inference. (performance & compatibility; waits on TEE remote attestation for AI workloads) [9][19]
- Safeguard attestation: Trust rests on a small number of hardware vendors, and a per-CPU Intel attestation key has been extracted by physical attack. (hardware trust; waits on TEE remote attestation for AI workloads) [7][9]
- Safeguard attestation: Safeguard evidence must be bound to the model actually served, which depends on model-identity attestation. (evidence binding; waits on Model identity attestation) [19][20]
- Safeguard attestation: No independent red-team or audit of a safeguard-attestation system has been published, and the available prototypes are described by their authors as proofs of concept that have not been stress-tested by a counterparty. (adversarial validation) [2][6]
- Bandwidth limits and compartmentalization: No prototype exists; RAND recommends prototyping key security features and integration now. (adversarial validation) [14]
- Bandwidth limits and compartmentalization: The report describes internal integrity checks, audit logging and accreditation, but no way for a party outside the operator to verify the facility's properties. (access & governance) [14]
- Bandwidth limits and compartmentalization: Human review of every prompt and response makes each request take three to five minutes, with the review steps as the rate-limiting factor. (performance & compatibility) [14]
- Bandwidth limits and compartmentalization: Detailed design information is withheld from the public report and is to be evaluated privately with stakeholders, which limits independent public scrutiny. (access & governance) [14]


## What the verifier sees

- Model weights: shown by none; depends on the design for Safeguard attestation; hidden by none; not involved in none; unspecified for Bandwidth limits and compartmentalization.
- Inputs and outputs: shown by none; depends on the design for Safeguard attestation; hidden by none; not involved in none; unspecified for Bandwidth limits and compartmentalization.
- Training data: shown by none; depends on the design for none; hidden by none; not involved in Safeguard attestation; unspecified for Bandwidth limits and compartmentalization.

## Implementations

- Safeguard attestation: none on the map
- 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. Proof-of-Guardrail in AI Agents and What (Not) to Trust from It, X. Jin et al. (2026). https://arxiv.org/abs/2603.05786
2. Verifiable-ClawGuard: proof-of-guardrail reference code, SaharaLabsAI (2026). https://github.com/SaharaLabsAI/Verifiable-ClawGuard
3. Safety Without Compromising on Privacy, D. McCann-Sayles et al. (2026). https://tinfoil.sh/blog/2026-09-14-safety-without-compromising-privacy
4. PAL*M: Property Attestation for Large Generative Models, P. Chantasantitam et al. (2026). https://arxiv.org/abs/2601.16199
5. Enabling Verifiably-Scoped Monitoring through Large Language Models and Trusted Compute, B. Penchas et al. (2026). https://icml.cc/virtual/2026/78630
6. Auditor-in-a-Box: Tools for Third-Party Auditing, R. Rinberg & B. Penchas (2026). https://www.lesswrong.com/posts/uWYk7MM9hAf9GEbGe/auditor-in-a-box-tools-for-third-party-auditing
7. TEE.fail: Breaking Trusted Execution Environments via DDR5 Memory Bus Interposition, J. Chuang et al. (2026). https://tee.fail/
8. DDRop: Active Memory Interposer Attacks on Confidential VMs by Dropping DDR5 Writes, J. De Meulemeester et al. (2026). https://ddropattack.eu/
9. On TEEs for Privacy-Preserving Monitoring in AI Governance, Gloria Z (2026). https://techgov.intelligence.org/blog/on-tees-for-privacy-preserving-monitoring-in-ai-governance
10. Battering RAM: Low-Cost Interposer Attacks on Confidential Computing via Dynamic Memory Aliasing, J. De Meulemeester et al. (2026). https://batteringram.eu/
11. RMPocalypse: How a Catch-22 Breaks AMD SEV-SNP, B. Schlüter & S. Shinde (2025). https://rmpocalypse.github.io/
12. SEV-SNP RMP Initialization Vulnerability (AMD-SB-3020), AMD (2025). https://www.amd.com/en/resources/product-security/bulletin/amd-sb-3020.html
13. 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
14. Highly Secure Inference Data Centers: A Vertically Integrated Strategy for Security Engineering, S. F. Comer et al. (2026). https://www.rand.org/pubs/research_reports/RRA4827-1.html
15. DiLoCo: Distributed Low-Communication Training of Language Models, A. Douillard et al. (2024). https://arxiv.org/abs/2311.08105
16. Does Distributed Training Undermine Compute Governance?, R. Rahman (2026). https://arxiv.org/abs/2605.29359
17. Traffic Shaping for Workload Classification, Lucid Computing (2026). https://lucidcomputing.substack.com/p/traffic-shaping-for-workload-classification
18. The Tray as a Bandwidth Boundary, Amodo Design (2026). https://amododesign.com/notes/2026-03-16-dpu-bandwidth-limiter/
19. Attestable Audits: Verifiable AI Safety Benchmarks Using Trusted Execution Environments, C. Schnabl et al. (2025). https://arxiv.org/abs/2506.23706
20. How Tinfoil Proves Exactly What Model Is Running, Tinfoil Team (2026). https://tinfoil.sh/blog/2026-02-03-proving-model-identity