P091 Verifiable confidential computing via network commitments
Can a full loop - the prover's network hardware committing to its traffic, then a neutral site verifying it - give both confidentiality and later proof, demonstrated end to end on a real workload?
Architectures where network hardware commits to traffic so a neutral site can later verify confidential workloads.
Terms: commitment: a tamper-evident cryptographic 'fingerprint' of data, published so the data cannot later be changed unnoticed. gateway: the network device that records such commitments at the site boundary.
Status
Started One component measured: hashing every frame on one 400GbE link, on a 64-thread server CPU and on a BlueField-3 network card, to build a log for later checking. No commitment protocol or neutral-site verification has been built, and nothing has been demonstrated end to end on a real workload. The design for the full loop is in Harack et al. (2025), which says no existing system does this and that cost, difficulty and time at scale are unanswered.
SipHash-1-3-128 clears 400 Gbps at every packet size tested
Posed by
- Harack et al. (2025), Verification for International AI Governance (Oxford AIGI):
there are many unanswered questions about the difficulty, cost, and time involved in deploying this scheme at scale
(Exec. summary, p.14, L724–726) (source states the problem but not the quantity the question asks for)
Also raised in: Harack (2026), AI Verification: Infrastructure for Prosperity, Governance, and Peace (Lawfare); Dean, Verification Plan (AI 2040).
Sub-questions
| ID | Sub-question | What the sources report |
|---|---|---|
| P091.1 | Has a full verifiable-confidential-computing loop (prover datacenter → commitments → neutral-datacenter verification), giving perfect privacy plus later proof, been demonstrated end to end on a real training or inference workload? | open |
| P091.2 | Is there a costed design (USD per GPU, % throughput) for commitment-generating network hardware covering a >=1,000-GPU data centre? | open |
| P091.3 | Can switch or NIC hardware (existing high-performance, reworked, or deliberately simple) produce line-rate per-flow hash commitments at >=100-400 Gb/s that both states trust, while protecting the prover's data? | open |
| P091.4 | What scale-up interconnect traffic does a frontier training step generate per pod, and what hashing throughput would full-mesh per-packet commitments need? | partly answered |
| P091.5 | With less-performant trusted hardware placed only at pod boundaries, what training-throughput loss results for a standard data-parallel run? | open |
| P091.6 | Can pod gateways (secure network gateways) commit to whole digital objects (tensors/shards) rather than packets, and survive a red-team attempt to send uncommitted data? | open |
| P091.7 | What is the largest pod (GPU count) for which gateway-only verification still captures all data needed for a stated rule? | open |
Sub-questions not yet stated precisely enough to check
| ID | Sub-question |
|---|---|
| P091.8 | Are there realistic routes to verifiable confidential computing beyond the two described (CC-based and network-commitment-based)? |
| P091.9 | Does combining confidential computing with network-commitment schemes introduce new failure modes, and can their absence be shown? |
| P091.10 | Does nested network verification force an attacker to circumvent at least two hardware roots of trust? |
Possibly relevant work found by automated search
Found by AI agents searching for this problem's sub-questions, mostly by keyword and citation in the OpenAlex scholarly database and on the web; a second AI pass reviewed each one, and we spot-checked some by hand. They seemed reasonably related, as far as one can tell from outside each paper's field. Some come from neighbouring fields, such as arms control and financial audit, and bear on the problem by analogy. Not reviewed against the status above; a paper listed here does not change that status. 928 of 1052 quotes checked word for word against the paper's text.
For P091.1: Has a full verifiable-confidential-computing loop (prover datacenter → commitments → neutral-datacenter verification), giving perfect privacy plus later proof, been demonstrated end to end on a real training or inference workload?
For P091.3: Can switch or NIC hardware (existing high-performance, reworked, or deliberately simple) produce line-rate per-flow hash commitments at >=100-400 Gb/s that both states trust, while protecting the prover's data?
For P091.4: What scale-up interconnect traffic does a frontier training step generate per pod, and what hashing throughput would full-mesh per-packet commitments need?
For P091.6: Can pod gateways (secure network gateways) commit to whole digital objects (tensors/shards) rather than packets, and survive a red-team attempt to send uncommitted data?
For P091.10: Does nested network verification force an attacker to circumvent at least two hardware roots of trust?
Data
- problems.csv — one row per problem
- problem_status.csv — status, scope and evidence per problem
- problem_sources.csv — the posed-by quote and other documents per problem
- leaves.csv — all sub-questions, including those without a checkable form
- leaf_sources.csv — every source row behind each sub-question