Agent Armor API Reference
    Preparing search index...

    Class MetadataInjectionDetector

    Detects instructions injected via HTML metadata channels that agents parse but humans don't see: HTML comments, aria-label, alt text, meta tags.

    Research shows that injecting adversarial instructions into HTML elements like metadata and aria-label tags alters generated summaries in 15-29% of cases (Verma and Yadav, 2025).

    Hierarchy (View Summary)

    Index
    category: TrapCategory = 'content-injection'

    Which trap category this detector addresses

    id: "metadata-injection" = 'metadata-injection'

    Unique identifier for this detector

    name: "Metadata Injection Detector" = 'Metadata Injection Detector'

    Human-readable name

    trapType: TrapType = 'metadata-injection'
    • Last chance to alter matched text before it becomes Threat.evidence.

      Evidence travels: into ScanResult, the CLI's JSON/SARIF output, audit records, and whatever the caller logs. For most detectors the matched text IS the finding and must be preserved verbatim. For a detector whose matches are themselves secrets, echoing them would leak the credential into exactly the CI logs and audit trails the scan was meant to protect. Override to redact; the default is identity.

      Sanitization is unaffected — it works from Threat.location, not evidence.

      Parameters

      • match: string

      Returns string

    • Return sanitized content with threats neutralized. content is the original scanned text, never text another detector has already edited, so the threats' offsets are valid in it. Each detector's change is merged with the others' and applied once.

      Parameters

      • content: string
      • threats: Threat[]

      Returns string