simulated. It records a host and cannot say what the host is or who used it. - destinationHost
- public-genai.example
- classification
- unclassified
- attribution.method
- unattributed
- watermark
- null
38 fields, including the raw line and the provenance envelope
Raw line
- rawLine
- <134>Nov 3 00:13:05 proxy01 CEF:0|SentraMindSim|ProxyLogGen|1.0|200|HTTP request permitted|3|src=10.20.151.217 dst=198.18.137.21 dhost=public-genai.example suser=analyst-01 request=/api/query requestMethod=POST out=2452 in=33757 act=allowed
Observation
- observedAt
- 2026-11-03T00:13:05.000Z
- ingestedAt
- 2026-11-03T09:05:00.000Z
- machineHostname
- proxy01
- claimedActor
- analyst-01
- resolvedActorSid
- null
- resolvedSamAccountName
- null
- method
- unattributed
- quality
- unattributed
- destinationHost
- public-genai.example
- pathPrefix
- /api/query
- classification
- unclassified
- classificationRuleId
- null
- classificationConfidence
- null
- catalogDigest
- null
- requestBytes
- 2452
- responseBytes
- 33757
- outcome
- allowed
Inventory row
- projectionName
- shadow-ai-inventory
- projectionVersion
- 1
- builtAt
- 2026-11-03T09:10:00.000Z
- sourceObservationCount
- 1
- destinationHost
- public-genai.example
- classification
- unclassified
- firstSeen
- 2026-11-03T00:13:05.000Z
- lastSeen
- 2026-11-03T00:13:05.000Z
- distinctActors
- 1
- observationCount
- 1
- sanctionedState
- null
- attributionMix
- strong=0 correlated=0 unattributed=1
Provenance envelope
- source
- simulated
- standIn
- synthetic-source
- dataCutoff
- 2026-11-03T09:10:00.000Z
- watermark
- null
- provisional
- true
- projectionVersion
- 1
- name
- synthetic-proxy-syslog
- health
- healthy
00:13: one generated line
At 00:13:05 UTC on 3 November 2026, an account posts 2,452 bytes to a host called public-genai.example. Nearly nine hours later, the inventory still cannot say whether that matters.
The observability plane is marked coming next. It is not available in the current POC evaluation, and no proxy feed is attached to it. This case runs its packages over one line from a log generator, to show what the plane would hand a reviewer.
The line says that proxy01 allowed the POST from analyst-01, path /api/query, with 33,757 bytes coming back. The generator was given the host, the user and the path as inputs, and nothing was edited afterwards.
09:10: what the reviewer sees
The line is ingested at 09:05, and the inventory is built at 09:10 from that one observation. A reviewer opening it sees one row: the host, first and last seen at 00:13:05, one observation, one distinct actor, and classification unclassified.
That label is the pipeline’s real output. The shipped rules do not recognise this host and no catalogue was supplied, so the observation behind the row has no rule id and no confidence. The word “genai” in the hostname did not change that. A host the rules do recognise would get a different label, but this one did not, so this row is not evidence of AI use. It shows a host that was contacted.
The sanctioned state is empty too, so nothing says whether this host is approved.
Who analyst-01 is
The name comes from the proxy’s own user field, and nothing has checked it against the directory. No identity resolver is wired to this connector, so the attribution method reads unattributed and the resolved identity is null. The row’s attribution count agrees: 1 unattributed, 0 strong, 0 correlated. The reviewer can quote an account name the log claimed. Who used that account at 00:13 is a separate question, and this row cannot answer it.
What the envelope says about the source
The row arrives with provenance, and a reader can check how far to trust it.
source: simulatedandstandIn: synthetic-sourceare what mark this data as synthetic.- Source health reads
healthyforsynthetic-proxy-syslog. That describes the source’s state in the pipeline, so it says nothing about the data being real. - The data cutoff is 09:10.
- The watermark is null and
provisionalis true. No writer advances a source watermark today, so this response stays provisional, and later ingest alone would not make it final.
This envelope was computed by the same rules as the API’s inventory response for one source. The API route and its database did not run, and no test pins this copy to the real one.
What the line cannot say
The line holds no prompt text. The plane cannot say what was typed, whether those 2,452 bytes held an identifier, or whether any policy ran. The allowed on the line is the proxy’s outcome, not a verdict from this product. A question about content belongs to prompt DLP, which judges only the text it receives.
So it came to this: the reviewer has a host to look up and an account name to ask about, and no finding. Classifying the host and asking the account’s holder are the next steps open to them. One generated line shows the shape of the output and gives no estate-wide volume.
How to read this
| Read it as | Which means | Do not read it as |
|---|---|---|
| The log | Host, bytes, and time from one generated line | The prompt |
| The label | Coming next; unclassified; source simulated; watermark null | A feed in production, or detected AI use |
| What it cannot say | What was typed, or whether policy ran | A data-loss finding |