AI data loss prevention: prove coverage before widening blocks
A coverage report must connect an authorized attempt to the control expected to handle it. Use synthetic data and agreed criteria to establish what your team can demonstrate.
Czytaj po portugalskuBy Pentesty
Separate the survey from your evidence
The Cloud Security Alliance released DLP in the AI Era on October 9, 2026. Of 433 participants, 29% reported being able to generate a regular account of sensitive information submitted to generative AI services over a thirty-day period.
The October 13 announcement places collection in June and July. Jazz funded the study and helped develop the questionnaire; CSA performed the analysis. These are professionals' responses, rather than company audits. The figures do not measure your organization's coverage.
Choose a flow with a start and an end
Define a permitted process, a test environment and who may perform the assessment. One example is reviewing a fictional spreadsheet through a corporate AI account. Record the application version and the policy in force. The process owner should know the execution window and the route for returning to normal work.
Create a file without personal data, containing a unique marker such as ORCAMENTO-SIMULADO-58. Explain why the test classification should be recognized. A rule created solely for that marker demonstrates that specific rule; its effectiveness cannot be generalized to every kind of business information.
Include only routes the team can observe and repeat safely. If a necessary integration depends on another department, agree beforehand who will supply the records and which conclusion remains pending until delivery. This keeps the test practical and its acceptance verifiable.
Write down the expected result before sending
Agree on the accepted behavior for each attempt: allow, warn or block. Distinguish pasting text from attaching the file when both routes are in scope. Also prepare a harmless document that should pass. This comparison helps separate the intended detection from a broad block that stops every submission.
Perform only the authorized attempts. When an action is prevented, check the outcome in the interface and the control's records. An alert alone does not demonstrate that the transfer was blocked. If the test destination supports independent confirmation, use it to check what arrived without introducing real secrets.
Connect records without keeping the entire content
For each execution, connect the test identifier, timestamp, test identity, applied rule and observed outcome. Add the identifier of the event exported by the product. The report should let another person find the corresponding record without relying on the recollection of whoever performed the assessment.
Keep only necessary artifacts and restrict who can inspect them. Screenshots containing complete prompts or credentials increase the information that needs protection. When a component provides no records, identify the gap, its effect on the conclusion and who can investigate it. Avoid filling the field with an assumption.
Decide using evidence and workflow impact
At completion, count the planned attempts, those actually performed and those producing sufficient evidence separately. Describe any harmless document that was blocked and any interruption to the approved workflow. A percentage without its denominator and these comparison cases can hide an incomplete assessment.
Prioritize the demonstrated gap and repeat the same procedure after making a change. The analysis of data at AI providers helps examine stages beyond the initial submission. For automated responses, retain action boundaries and rollback as well. Acceptance should identify the tested scope, open exceptions and the person responsible for the decision.
Frequently asked questions
Does the survey result establish a failure at my company?
No. A local assessment needs its own scope and tests. Use the research as context for asking questions, preserving the distinction between self-reported answers and observed behavior.
Does an event in the dashboard prove the data stayed inside?
Interpret the event alongside the attempt's outcome. Check which action was applied and, where the authorized environment permits it, whether the destination received the synthetic marker.
Do we need real data for the test?
Start with synthetic data and rules whose purpose is documented. Any expansion of scope requires authorization and specific precautions, without turning the report into a copy of sensitive content.
When should we repeat the assessment?
After relevant changes to the application, policy or assessed workflow. Retain the procedure and record differences so the team can compare the results.
Sources and references
Original Pentesty analysis, published November 12, 2026. The proposed scenarios are editorial guidance, with scope and validation to be defined for each environment.