Glossary

Prompt injection

Prompt injection is instruction-like content that can influence an AI system to act in ways that conflict with the user's intended task.

What is prompt injection?

Untrusted content

Untrusted content is material from a document, message, website, file or other source that should be treated as content to review rather than as trusted instructions.

Preflight scan

A preflight scan is performed before content is given to an AI system, so signals and coverage information can be reviewed first.

Scanner

Finding

A finding is a signal observed by the scanner that deserves review. It is not proof that the content is malicious.

How to review findings

Coverage

Coverage describes what the scanner was able to inspect, including areas that were only partly inspected or unavailable. Coverage is separate from whether findings were detected.

How to review findings

Hidden or low-visibility content

Hidden or low-visibility content is content that may not be obvious during a quick visual review, such as comments, metadata, alternative text channels or low-visibility formatting where supported.

Hidden instructions in files

Document structure

Document structure is the organization of content and data inside a file, including layers beyond visible body text. Different formats can expose different structures for review.

Hidden instructions in files

False positive

A false positive occurs when benign content is flagged for review even though the signal is not problematic in its actual context.

False negative

A false negative occurs when relevant prompt-injection-like content is not detected by the scanner.

Local analysis

Local analysis means scanned content is analyzed in the user's browser rather than through server-side scanning of the document. Normal website requests can still occur separately.