P061 Verifiable memory wipes and memory challenges
Can a node's HBM/DRAM/SSD and network-equipment memory be proven wiped or filled within a stated time and challenge latency, and how much non-wipeable storage remains?
Proving HBM/DRAM/SSD and network-equipment memory is wiped or filled: wipe time, non-wipeable storage, challenge latency and soundness.
Status
Started Proof-of-secure-erasure experiments on H200 hardware, the authors' own report: RAM and HBM wipe estimated at tens of minutes for a GB200 tray; disk wipe about 24 hours for 15 TB, extrapolated from a measured 1 TB NVMe write; challenge-phase time not yet computed; non-wipeable memory targeted below 1 GiB but not yet quantified; network-switch memory listed as open. Challenge clause: Monfared et al. 2026 measured a gap above 350 ms between HBM-resident and host-RAM-resident 60 GB challenges on an H100, with no false-alarm rates; Reuter 2026's roughly 2x margin against host-DRAM caching is an analysis; Bursuc et al. 2024 give provably secure proofs of erasure with round-trip-time bounds (theory). Cankaya's v0.2 draft (dated) says a network-level timing probe telling one node's DRAM from another's 'has not yet been demonstrated'. Both clauses sit in one memory-verification community; not every clause answered.
Writing a full 1 TB gave a final average write speed of 160 MiB/s — much lower than the advertised rate — taking ~1h 35 minutes.
Posed by
- Cankaya (2026), A System Overview for Near-Term, Low-Trust AI Compute Verification (v0.2):
To my knowledge, distinguishing content of DRAM (HBM, DDR) between one server node and another, using a challenge+response timing probe on any network level, has not yet been demonstrated.
(5.1.2 Memory challenging and memory wiping - Open research questions; LessWrong v0.2 (rev. 1.4.0, 2026-06-24), L463 of the archived text) (source states the problem but not the quantity the question asks for)
Also raised in: AI 2040 (2026), Get Involved in Verification; AI Security Forum, Requests for Discussion; Amodo (2026), AI 2040 Plan A — Verification SITREP; Dean, Verification Plan (AI 2040); Proofworks, Living Verification Documents; Reynolds & Foster (2026), Network Traffic Hashing; Reynolds & Jacobi (2026), Memory Wipes - Performance Analysis.
Sub-questions
| ID | Sub-question | What the sources report |
|---|---|---|
| P061.1 | Can a cluster node's HBM + NVMe be provably and periodically wiped quickly, at what cost and frequency? | partly answered |
| P061.2 | How long does wiping a rack's (e.g. NVL72) disks take, given the disk-write ceiling? | partly answered |
| P061.3 | How many bytes of non-wipeable storage (firmware stores, unpublished component memory) does an H100 server / GB200 tray contain under a minimal image? | open |
| P061.4 | Can a rack-wide proof of secure erasure with synchronised challenges run on an NVL72 within the RTT bound? | open |
| P061.5 | How long does the challenge phase take for a GB200 tray at a stated soundness level? | open |
| P061.6 | Can a cheating prover recompute memory-filler labels (e.g. SHA-iterated filler) with a co-located accelerator faster than the challenge window / ~1 ms RTT? | implied by a source, not stated as open |
| P061.7 | Can host-DRAM caching or NVMe overflow of evicted state meet a stated challenge-latency bound, and can a pinger distinguish it from an HBM read? | partly answered |
| P061.8 | What throughput loss results from filling all free HBM with verifiable filler? | open |
| P061.9 | Can a network-side timing probe distinguish local DRAM from RDMA-remote DRAM (median/p99 latency gap vs jitter at each probe position), and how close must the probe sit? | open |
| P061.10 | How much hidden capacity remains on enterprise SSDs after a full verifiable wipe (over-provisioned NAND)? | open |
Sub-questions not yet stated precisely enough to check
| ID | Sub-question |
|---|---|
| P061.11 | How should memory challenges handle dynamic memory allocation? |
| P061.12 | Are partial or cumulative wipes sound? |
| P061.13 | Can GPU drivers be included in the minimal wipe image without breaking its guarantees? |
| P061.14 | Can memory-challenge verification prove completeness? |
| P061.15 | Can hardware backdoors inside the auditor's redline survive reboot and memory wipes? |
| P061.16 | What data-centre architectures allow storage to be separated from compute trays? |
| P061.17 | Can per-link traffic hashing scale to 10k+ links without a ~2U server per link? |
| P061.18 | How can network switches (frontend, backend, management) have their memory verifiably wiped? |
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 P061.1: Can a cluster node's HBM + NVMe be provably and periodically wiped quickly, at what cost and frequency?
For P061.2: How long does wiping a rack's (e.g. NVL72) disks take, given the disk-write ceiling?
For P061.3: How many bytes of non-wipeable storage (firmware stores, unpublished component memory) does an H100 server / GB200 tray contain under a minimal image?
For P061.4: Can a rack-wide proof of secure erasure with synchronised challenges run on an NVL72 within the RTT bound?
For P061.6: Can a cheating prover recompute memory-filler labels (e.g. SHA-iterated filler) with a co-located accelerator faster than the challenge window / ~1 ms RTT?
* 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