You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
{{ message }}
Repository navigation
Commit b6687a1
Browse filesBrowse the repository at this point in the historyBrowse files
Copy file name to clipboardExpand all lines: THIRD-PARTY.md
+1-1Lines changed: 1 addition & 1 deletion
Display the source diff
Display the rich diff
Original file line number
Diff line number
Diff line change
@@ -33,7 +33,7 @@ theirs to claim and ours to relay, clearly labelled as relayed.
33
33
|**AWS CloudTrail** (digest files) |**L0**, blocked only by **VLC-L1-3**| A real chain: `previousDigestHashValue` over the previous digest **file**, RSA signature in S3 object metadata, and an hour with no activity still emits a digest with `logFiles: []` — a rare explicit *"nothing this interval"* assertion | No end marker. Truncating the newest digest is undetectable **from the chain alone**; you need out-of-band knowledge of which hour it should be. Chain is over files, not events; no dropped-event count exists in the product; scope is in `EventSelector` config, not in the evidence. |
34
34
|**MCP tool calls** (`tools/call` over JSON-RPC 2.0) |**L0**| a request/response correlation `id`; a `tools/list` result that is genuine coverage information | The `id` correlates a call with its reply; it is not an ordinal over the delivered file, and hosts assign it per connection, so absent calls leave no hole. No integrity field, no record counter, no requirement to state which servers were connected. NSA's June 2026 MCP information sheet names *"poor or missing audit logs"* as a risk and recommends logging every invocation — and specifies no mechanism for integrity, loss or coverage. |
35
35
|**Fab equipment** (SECS-II/GEM over HSMS, SEMI E5/E30/E37) — *scored from public documentation, not from the paywalled standards; see the adapter's `_basis`*|**L0**|`SMPLN`, a real ordinal over the trace samples of one `TRID`; `S2F23`, which names the SVIDs a trace requested | No per-message digest appears in the public documentation or the open-source implementations, and HSMS is a session protocol over TCP/IP — it protects the hop, not the stored row. The ordinal covers trace samples only, not event reports or alarms. The larger loss needs no standard to establish: the **sampling period itself** is undeclared dead time — at one sample a minute, fifty-nine seconds in sixty go unobserved and no record marks them as unobserved rather than uneventful. Which SVIDs were enabled, and which wafers the sampling plan measured, live in equipment configuration. **Each of these is written in the adapter as a question a clause reference would settle.**|
36
-
|**NVIDIA OpenShell** 0.1.2 (Open Agent Safety Platform), the sandbox event stream held by the gateway — *live capture, 2026-09-29; see the adapter's `_basis`*|**L0**, blocked by **VLC-L1-1**| every denial carries a reason; every allow names the policy rule that decided it; policy loads and reloads carry the policy's version and full SHA-256; the record is held **outside the agent** (VLC-L5-1 passes) | No integrity field, no sequence number, no drop accounting, and the lines are not in timestamp order (7 of 75 adjacent pairs run backwards). A decision line does not carry the hash of the policy it was decided under. Removing the six DENIED lines leaves a log that scores identically. **Corrected from the documentation-based score of 2026-09-29**, which assumed the agent could write its own log: in the capture it could not, and the OCSF JSON file the documentation describes was not produced at all on this driver. **Sentry is not scored.**|
36
+
| **NVIDIA OpenShell** 0.1.2 (Open Agent Safety Platform), the sandbox event stream held by the gateway — *live capture, 2026-09-29; see the adapter's `_basis`* | **L0**, blocked by **VLC-L1-1** | every denial carries a reason; every allow names the policy rule that decided it; policy loads and reloads carry the policy's version and full SHA-256; the record is held **outside the agent** (VLC-L5-1 passes) | No integrity field, no sequence number, no drop accounting, and the lines are not in timestamp order (7 of 75 adjacent pairs run backwards). A decision line does not carry the hash of the policy it was decided under. Removing the six DENIED lines leaves a log that scores identically. **Corrected from the documentation-based score of 2026-09-29**, which assumed the agent could write its own log: in the capture it could not, and the OCSF JSON file the documentation describes was not produced at all on this driver. **Sentry is not scored.** Reported upstream with a proposed fix: [NVIDIA/OpenShell#3817](https://github.com/NVIDIA/OpenShell/issues/3817). |
37
37
|**Agent tool-call transcripts** (every framework) |**L0**| readability | Written by the audited process. Fails **VLC-L5-1** by construction, whatever else is done to them. |
38
38
|**JIDEC ledger, before 2026-09-28** (HORIZON SHIELD), *scored by its own operator; kept as a control*|**L0**, blocked by **VLC-L1-1** and **VLC-L1-3**| a real ordinal `n`; SHA-256 over each entry's canonical bytes; every hash anchored individually to Bitcoin through OpenTimestamps | No predecessor binding: entries verify one at a time, so removing one or truncating the newest leaves every survivor valid. A gap in `n` is visible; the tail is not, because nothing states how many entries exist and there is no end marker. The Bitcoin anchor proves a hash existed by a block time, not that the ledger delivered every hash it anchored. The entries are written by the operator whose own work several of them record. |
39
39
|**JIDEC ledger** (HORIZON SHIELD), *live export, scored by its own operator after jidec-chain-v1*|**L1**; blocked from L2 by **VLC-L2-1** and the rest of L2 |`prev_entry_sha256` and `entry_sha256` on every row by this specification's `sha256-canonical-fields` recipe, derived from fields each entry already carried; a `jidec-head-v1` end marker; the head also written into each daily batch whose bytes are stamped to Bitcoin | No loss accounting: nothing states how many submissions arrived or were dropped, only what was appended. Coverage and policy unchanged. On the log alone L1 is internal consistency; a complete rewrite shows only against a held head, which is what the stamped batch provides. |
0 commit comments