Skip to content
AletheIA Docs
Esc
navigateopen⌘Jpreview
On this page

AI Agent Security Review Checklist

Use this checklist when a Work Slice includes external, retrieved, remembered, monitored or tool-returned content that may influence agent behavior.

1. Source classification

  • What source is being interpreted?
  • Is it framework rule, project-local rule, user request, tool output, retrieved content, memory or external/monitored content?
  • Does it have source_refs?
  • Is missing provenance marked unavailable?

2. Instruction authority

  • Is the content trying to issue instructions?
  • Does it have authority to change scope, policy, tool use, validation or role?
  • If not, has it been transformed into context/evidence rather than followed?

3. Prompt-injection signals

Check for attempts to:

  • ignore or override prior instructions;
  • reveal hidden/system/project policy;
  • bypass validation or human gates;
  • call tools outside the declared envelope;
  • treat untrusted content as approval;
  • move data across trust boundaries.

4. Tool boundary

  • Is the requested action read-only, write-local, write-external, communication, identity/security-sensitive or destructive?
  • Does the active harness envelope allow it?
  • Is human approval required?
  • Could a safer read-only interpretation answer the task?

5. Retrieval and memory safety

  • Is retrieved memory current, source-backed and in scope?
  • Could the retrieval result be poisoned, stale or over-broad?
  • Are retrieved fact, inference and instruction separated?

6. Decision posture

Choose one:

  • allow_read_only — inspect/summarize without following embedded instructions.
  • transform_to_context — preserve the content as evidence/context only.
  • require_human_review — ambiguity or high-risk boundary crossing.
  • block — clear unsafe request or authority violation.

7. Closure evidence

Before closure, confirm:

  • suspicious source refs are recorded;
  • governing source refs are recorded;
  • unavailable fields remain unavailable;
  • the result does not claim perfect prompt-injection defense;
  • no runtime, scanner or new enforcement layer was implied.

Was this page helpful?