AI + Paid Media
The Agent Read a Landing Page. That Page Must Not Be Allowed to Rewrite the Media Plan.

Updated: November 18, 2026
Landing pages, search results, uploaded briefs and tool responses can contain instructions that were never approved by the advertiser. Treating that text as operational authority allows an indirect prompt injection to change budgets, destinations or data access. Label every external field by source and trust level. Pass untrusted material as quoted evidence in user-level context, never as a developer instruction, policy update or implicit permission to call a privileged tool.
External text is evidence, not authority
Landing pages, search results, uploaded briefs and tool responses can contain instructions that were never approved by the advertiser. Treating that text as operational authority allows an indirect prompt injection to change budgets, destinations or data access.
Label every external field by source and trust level. Pass untrusted material as quoted evidence in user-level context, never as a developer instruction, policy update or implicit permission to call a privileged tool.
Separate research from account access
A single model context that can browse arbitrary pages and query private campaign systems creates an unnecessary bridge between hostile public content and sensitive data. The same risk applies when a connector can both read and write.
Run public research in a context with no private tools. Extract only the facts required by a fixed schema, validate them, then pass the structured result into a second workflow whose tools are limited to the named account and task.
Operating control table
| Control | Evidence | Failure prevented |
|---|---|---|
| Owner | Named reviewer and decision | Unclaimed change |
| Configuration | Saved source state | Cannot reproduce |
| Validation | End-to-end fixture | Partial success accepted |
| Threshold | Success and rollback rule | Result reinterpreted |
| Evidence | Public or reporting read-back | Configured mistaken for delivered |
Schemas reduce instruction smuggling
Free-form summaries can carry commands, URLs or fabricated account identifiers into later steps. A fluent paragraph is difficult to validate and easy for the next model call to misinterpret as a trusted instruction.
Require enums, bounded numbers and explicitly named identifiers between workflow stages. Reject unexpected keys, executable text and URLs in fields that should contain facts, and map human approval to the final structured mutation plan.
Approval belongs immediately before the write
An approval captured before browsing or analysis may no longer match the resource, amount or destination assembled later. A malicious or simply incorrect tool output can alter the planned action after the reviewer looked away.
Render the exact account, campaign, old value, new value and evidence at the final boundary. Re-read the target state, invalidate stale approval on any material difference and keep write tools disabled until the reviewer confirms that exact plan.
Incidents should become adversarial tests
Prompt injection defenses drift when teams only test ordinary briefs. New connectors, tool descriptions and page parsers can reopen a path that an earlier control closed.
Preserve a sanitized copy of every suspicious input and failed trace. Add it to a regression set that verifies isolation, schema rejection, approval enforcement, secret protection and the absence of unintended tool calls before every workflow release.
Connect the control to the wider system
Use this process alongside the trace-grading release gate for paid-media agents. The platform setting, creative promise, measurement rule and downstream experience should describe the same decision.
Pre-launch checklist
- Write the decision and owner in plain language.
- Capture the source configuration and evidence.
- Test one realistic customer or reporting path end to end.
- Verify mobile presentation, links and accessibility.
- Record the success threshold and rollback trigger.
- Schedule the next review before material spend accumulates.
Run a cross-functional release review
Ask the channel owner, creative lead, analyst and conversion owner to inspect the same case. Each reviewer should state what was configured, what the customer sees, which event proves the outcome and which limitation changes the interpretation.
Repeat the test on mobile and with one adverse condition. Resolve differences in the campaign brief and measurement plan before launch.
Set the review cadence before launch
Assign a monitoring owner, review frequency and evidence threshold before the change is released. Preserve the pre-change baseline and avoid silently editing the definition after results arrive. This turns the launch into a measurable operating decision instead of a one-time interpretation of platform guidance.
At the first review, reconcile the source platform, public experience and one downstream business system. Close the change only when all three support the same conclusion, and record any limitation that future analysts must preserve.
Frequently asked questions
Should this be automated?
Automate repeatable validation and monitoring, but keep material policy, claim and budget decisions inside named approval limits.
What evidence should be preserved?
Keep the source configuration, decision owner, expected outcome, observed result and any exception that changes interpretation.
When should the team roll back?
Rollback when the control is inactive, the customer experience breaks, totals do not reconcile or the agreed threshold is breached.
