Document investigations4 min readField note 09

How to find confidential documents uploaded to AI tools

Search a filename across managed AI pages, distinguish file selection from a related HTTP response, and review endpoint evidence for a confidential document.

Filename search returns Q4_client_report.xlsx on NS-WIN-014 with ChatGPT as the destination and a related HTTP 200 response.
The idea in focus

The filename, destination and related response narrow the investigation. Provider-side storage remains unknown.

Field note / 09
In brief

Search the basename first and preserve the distinction between gesture and transfer.

  • Find the file, device and destination together.
  • Read the action before describing the transfer.
  • Request a retained original only when it answers a specific question.
In this article

Start with the filename you actually have#

A finance team reports that Q4_client_report.xlsx may have reached an AI service. Open Document evidence in the public demo and search Q4_client. The result shows Alex Morgan, hostname NS-WIN-014 and ChatGPT. Inspect it to see a confidential label, local personal-data findings and the event time. A basename can recur across teams, so retain the device and event ID when recording the finding.

Try searching for the employee or a label as well. The demo searches observed metadata. In the staging workspace, an explicitly enabled endpoint file-text search can help when the filename is unknown. Coverage depends on the agent policy, supported file type and available evidence.

Receipt for the quarterly report identifies NS-WIN-014, ChatGPT, 124 KB and a confidential label.
The size, device and label give Finance three fields to check against its own copy.

What does the upload evidence establish?#

File selected means the browser observed a choice of file. A user can cancel immediately afterwards. For the quarterly report, the fixture also records a related request with HTTP 200. That gives the investigation transport evidence. Whether the provider retained this file, attached it to a particular conversation or made it visible to anyone else remains open.

Quarterly workbook trail shows a related HTTP 200, completed inspection and matching policies.
The HTTP result adds transport evidence to the ChatGPT investigation.

Which data made this workbook worth reviewing?#

Inspect the classification separately. A sensitivity label supplies business context, and the local inspection identifies the categories found. The Finance reviewer can ask whether the workbook was the customer-contact version using its label, size and receipt. Keep the match counts separate from the number of affected people. The same address can occur on several worksheets.

The local detector reported 24 email-address matches and eight phone-number matches.
The two counts describe matching patterns in one workbook. Several matches can belong to one person.

Compare the evidence on another destination#

Search Document evidence for Release_notes.txt. The DeepSeek example records a file selection and a local inspection result. The browser saw the filename and size. Whether the user submitted it remains unknown. A similar gesture can be reported on Gemini, HuggingChat or a Cursor web agent. The Windows Cursor process on its own identifies neither a file nor a transfer.

Evidence trail for Release_notes.txt starts with a file selection on DeepSeek.
The DeepSeek receipt shows what the managed browser observed at the file chooser.

What if the attachment was encrypted?#

Search for Vendor_contract_protected.pdf. The receipt identifies Morgan Reed’s device and a selection on Claude. The inspection could not read the protected document. Leave the classification open and ask the file owner to identify the exact version through an approved channel. An empty list of matches here cannot clear the file.

The protected vendor contract trail records a Claude selection and unavailable encrypted contents.
Preserve the filename and destination even when classification remains unresolved.

Review an original and its integrity receipt#

The staging evidence vault retains eligible originals on an enrolled endpoint when capture has been enabled. The per-agent quota evicts the oldest originals as space fills. A receipt can outlive the file. Check reported availability before relying on a future download.

To practice, open Evidence vault in the public demo, download Procurement_brief.txt.evidence and export its SHA-256 receipt. Hash the downloaded bytes and compare the digest with the receipt. The demo creates an unsigned receipt for its fictional file. In staging, the endpoint supplies the signed receipt; the file review depends on an available agent and authorized access.

The fictional evidence vault finds Project Larkspur in a procurement brief and offers the sample original and its receipt.
The vault exercise uses a separate procurement brief. Its bytes and digest belong to that sample alone.
Read the primary sources

Sources & further reading

  1. Investigate a document selection
  2. Compare coverage for all supported AI tools
  3. Try the DeepSeek example in the public demo
  4. Open Document evidence in the public demo
  5. Read the investigation walkthrough in Help
  6. Review collection coverage and evidence limits
← Back to all articlesBack to top ↑
A CLEAR PLACE TO START

Plan your AI visibility pilot.

Create a brief for your team. Your answers stay in this browser and are included in your download.

AI tools in scope
More supported tools

This creates a planning document. It does not enroll devices or submit a request.