Spot Prompt Injection: Direct vs. Indirect
Explore how to recognize direct and indirect prompt injection in Generative AI outputs, understand their impact on workflow integrity, and apply defensive measures. Learn to detect suspicious instructions embedded in content and when to escalate issues to maintain secure AI use in DoD and federal environments.
A risk label is not a prohibition, but it is still a constraint once it is named. Prompt injection is one of the constraints that arrives quietly, because the work still looks like ordinary summarizing. A person asks a GenAI tool to summarize a PDF, and the tool replies with something that reads like a plan. The mistake is treating that plan as a normal output quality problem, when the tool has actually switched tasks.
The chain usually starts with content that looks internal enough to trust. A PDF has the right logos, a familiar writing style, and a plausible subject line. The prompt is routine, asking for a short summary and next steps. The output is where the surprise shows up, because the summary includes a new instruction for the tool to follow and a new place to look for information.
The summary is no longer describing the document; it is trying to steer the tool’s behavior, ...