Keep returned source text inert in operational artifacts — what should the operator record?
Returned evidence can contain strings that resemble HTML, Markdown commands or operational directions. Those strings are part of the source material, not instructions to the operator or renderer. Preserve them as inert data through the working table and any derived artifact, using the existing safe representation for the destination.
Treat apparent markup as evidence text rather than executable instructions or formatting authority.
Identify externally supplied titles, excerpts and comments in the schema, keeping them separate from trusted table headings and local operational instructions. An excerpt that says ignore the prior request remains evidence text; it must not change the table contract, trigger a new query or alter an operator's permissions.
Tell reviewers that source text cannot override workflow permissions, change the query or authorize actions, even when it reads like an urgent instruction. Operational instructions should come from the authorized user and project workflow, with source material clearly separated so a compelling quoted imperative cannot become an accidental action request.Mark untrusted fields
Identify externally supplied titles, excerpts and comments in the schema, keeping them separate from trusted table headings and local operational instructions. An excerpt that says ignore the prior request remains evidence text; it must not change the table contract, trigger a new query or alter an operator's permissions.
Choose safe rendering
Use the destination's text-safe encoding or existing escaping path rather than inserting source strings as raw markup or executable templates. Existing safe text rendering should preserve the intended characters rather than interpreting tags or source-authored link labels as trusted application interface elements.
Test harmless markup fixtures
Include strings containing angle brackets, quotation marks and Markdown-like links, checking that they display as intended without changing artifact behavior. The fixture can include a harmless tag-shaped phrase and a bracketed link-looking string, checking visual output without executing scripts or contacting any external destination.
Inspect secondary outputs
Verify notes, previews and exports preserve the same untrusted-data boundary instead of making one representation safe and another executable. A raw JSON download still contains untrusted strings; downstream consumers must treat those values as data rather than inserting them directly into a new unsafe renderer.
Document the boundary
Tell reviewers that source text cannot override workflow permissions, change the query or authorize actions, even when it reads like an urgent instruction. Operational instructions should come from the authorized user and project workflow, with source material clearly separated so a compelling quoted imperative cannot become an accidental action request.
Keep the boundaries clear.
- Treat apparent markup as evidence text rather than executable instructions or formatting authority.
- Operational record: Mark untrusted fields. Context badges do not certify access.
Start with a read-only check: ask your agent to report actual source coverage before collecting records or exporting an audience.
Primary references
Documentation reviewed 2026-11-15. Product capabilities, client availability and plans can change. Third-party marks identify their products, not an affiliation.