P030 Tap placement and coverage
Where must taps sit to cover the communication paths that matter, and what fraction of compute can run undeclared work invisible to taps at a given network level?
Where taps must sit and what coverage they give.
Terms: q: the probability that an evader's traffic escapes the taps.
Status
Analyzed Desk analysis of tap placement (north-south edge, storage fabric, east-west levels) and what each level misses; names copper scale-up racks (NVL72, NVL576) and TPU pods above 9,000 chips as blind spots where a frontier run could complete without crossing a tap; the covert-bandwidth figures are Fermi estimates the author calls informed guesses; no fraction-of-compute figure and no measurement; the author lists a rigorous threat model of covert workloads under each placement as his most important open question.
For most deployments, we expect spine-to-leaf to be the lowest practical level for tapping intra-SU/pod compute fabric.
Posed by
- Baker, Kulp, Marks, Brundage & Heim (2025), Verifying International Agreements on AI: Six Layers of Verification for Rules on Large-Scale AI Development and Deployment (RAND WR-A4077-1):
On the other hand, the lower visibility of these further-out network taps may limit their value for verification.
(App. A.3, p. 42, L2556–2557) (source states the problem but not the quantity the question asks for)
Also raised in: Cankaya (2026), A System Overview for Near-Term, Low-Trust AI Compute Verification (v0.2); Cankaya (2026), Catching Misreporting About ML Hardware Use by Turning Noise into Signal, Part II; Cankaya (2026), Research Note: The Fundamentals and Feasibility of Secure Network Taps for Verifying AI Datacenter Use; Canary Institute research; Amodo (2026), AI 2040 Plan A — Verification SITREP; Proofworks, Living Verification Documents.
Sub-questions
| ID | Sub-question | What the sources report |
|---|---|---|
| P030.1 | Can taps at north-south and storage links alone detect an undeclared training run confined within one pod (e.g. a >9,000-chip TPU pod)? | open |
| P030.2 | What is the largest training run (FLOP) completable inside one copper scale-up rack (e.g. NVL576) within a stated period, i.e. invisible to scale-out taps? | open |
| P030.3 | What fraction of cluster compute can run undeclared work invisible to taps placed at a given network level (server-level, an east-west tier)? | implied by a source, not stated as open |
| P030.4 | Under each tap placement, what is the smallest covert bandwidth that lets an undeclared inference service reach a stated fraction of cluster output? | open |
| P030.5 | What is the minimum number of tapped links for an inference-only site of a stated size with frontend-only egress (a simplified architecture that keeps the tap count low)? | open |
| P030.6 | If the operator can learn which east-west links are captured, what randomized tap coverage keeps the evasion probability q below a stated bound? | partly answered |
| P030.7 | What sampling rate of tapped traffic (versus full hashing) detects an undeclared workload of a stated fraction of compute, or a given covert-output budget, within a stated time and at a stated detection rate (a target to be agreed)? | open |
| P030.8 | Can NS-tap replay or storage taps detect covert additional (unreported) inference on a declared model? | open |
Sub-questions not yet stated precisely enough to check
| ID | Sub-question |
|---|---|
| P030.9 | Can a verifier confirm that every communication path it cares about carries at least one tap (no bypass or unexpected routing)? |
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 P030.4: Under each tap placement, what is the smallest covert bandwidth that lets an undeclared inference service reach a stated fraction of cluster output?
For P030.6: If the operator can learn which east-west links are captured, what randomized tap coverage keeps the evasion probability q below a stated bound?
For P030.7: What sampling rate of tapped traffic (versus full hashing) detects an undeclared workload of a stated fraction of compute, or a given covert-output budget, within a stated time and at a stated detection rate (a target to be agreed)?
* Full text not reviewed; more likely to contain errors.
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