Indirect Prompt Injection Scanning

Indirect prompt injection hides instructions inside content an agent later reads — a document, an attachment, a retrieved page. EVE scans ingested artifacts before they reach the model or agent, using deterministic content scanners, and blocks, quarantines, or requires approval — with signed scan evidence bound to the exact bytes scanned.

How scanning works

  • Real type detection — EVE detects the true media type from content, so a payload disguised as a benign file is parsed as what it actually is.
  • Bounded parsing + deterministic scanners — parsers run under safe-parsing limits and pattern/structure scanners produce a deterministic disposition before ingestion or use.
  • Digest-bound evidence — the scan evidence binds the artifact digest to the decision, so swapping the artifact after approval is detected at verification.
  • Honest completeness — incomplete extraction is reported as partial and cannot be presented as complete.

Evidence example

scan evidence { final_action: BLOCK, findings_digest, completeness_status: "partial" }

Readiness

Deterministic artifact/content scanning with signed evidence is PILOT_READY; artifact governance is customer-pilot ready at supported boundaries (Python; embedded-service and sidecar modes), covered by the artifact-scan phase report (24 tests).

Limitations

  • Scanners are deterministic and pattern/structure based; they detect what they are configured to match and are not a general-purpose understanding of arbitrary content.
  • OCR, vision, and QR/barcode reading are probabilistic and disabled by default (recorded as unavailable), so text hidden only in images may not be extracted unless those scanners are enabled.
  • Digest binding detects post-approval substitution at verification time; enforcing re-scan-before-use is a wiring responsibility at the ingestion boundary.

Next step

Run your RAG source documents through EVE scanning in a pilot and review the signed dispositions before they reach your agents.

Part of the EVE AI Core control plane Deterministic AI Governance Control Plane → Policy decisions that return the same result for the same input every time, before execution.