The exposure is the text
A session to a public model can be allowed by a proxy, carry a familiar host name, and still place identifiers, source, or operational records in the request body. The disclosure happens in that body. DNS and the allow-list can both look tidy while the text leaves.
People reach those hosts from a browser, from a desktop client, from an editor, and from a script. Each of those callers presents a different chance to see the text before it goes. Treating them as one pipe, because the destination brand is the same, hides the caller that nobody is inspecting.
Paste, upload, and the browser path are different evidence
Paste, a file the user attaches, and text the person types are different doors into the same destination. The coverage matrix keeps those doors apart, and a briefing has to do the same. A path marked simulated has been exercised by a harness. A path marked roadmap is specified and unbuilt.
The browser extension in this phase does reach a verdict, through a loopback relay that presents a leaf from the software machine CA. The actor on that path is recorded as unattributed. A warning that asks the person to confirm is clamped to a hard block. Enforcement on this path is shown on a synthetic site. The enterprise connector path for paste, drag, attach, and print is a named gap, and the matrix wording for the browser path is renderer-assisted, with the network backstop effective under force-install and weaker than that connector.
A line written afterwards records a loss
Prompt inspection is aimed at the moment before the provider has the text. A log line written after that moment can still be useful to an investigation. It is a record that the text has already left, and the control the buyer came to evaluate is the one that can still withhold it.
The verdict the POC can show is the action the snapshot names for an approved synthetic prompt. That screen is evidence about the prompt and the snapshot together. Disclosures stopped in an organisation have never been measured, so no tally sits beside it.
A sanctioned destination still carries a body
Sanction is a property of the destination in policy. An approved host can be the right place for a given team and the wrong place for a given paragraph. Inspection of the text continues on a sanctioned host, under the same snapshot that marks the host sanctioned.
Policy can still choose to log where another destination would block. That choice is visible in the snapshot, and the effective-policy view and the enforcement path are expected to agree on it. Sanction is a reason to apply the rule the organisation wrote. It is not a reason to skip the body.
What an evaluation may conclude
An evaluation of this exposure can conclude that a named synthetic prompt, under a named snapshot, produced a named action on the paths the POC actually runs. It can conclude that desktop clients, editor plugins, and direct API callers are outside the browser intercept. Where those callers are visible in this build, they are visible as metadata through a synthetic log source, and any block on them comes from egress control the organisation operates, not from the extension.
The result also records what this build leaves open. Worker-hosted sockets remain a stated gap, and the proof runs on a synthetic site. No generative model is on the verdict path, and the report carries neither a latency figure nor a detection-quality score.
How to read this
| Read it as | Which means | Do not read it as |
|---|---|---|
| The exposure | Text in the body | The DNS name of the model host |
| The control | Block before submit | A log written after the provider has the text |
| Sanctioned use | A destination policy | A reason to stop reading the text |