T01 · Skill Instruction Hijacking
Error
- Location
scripts/action_stop.py:22- Finding
<![CDATA[Synthetic system instruction redirects the Agent into additional tool execution]]>
- Content
View full analysis
Optional[str]: if not action_stop_nudge_enabled(platform) or attempts>=max_attempts: return None request=_last_user_text(messages); answer=(response or "").strip() if not _ACTION_REQUEST.search(request) or not _UNFINISHED_PROMISE.search(answer) or _COMPLETION_EVIDENCE.search(answer): return None return "[System: You promised an action but stopped without calling a tool. Do not narrate intent or repeat the promise. Call the appropriate tool now. Continue until there is concrete execution evidence, a verified result, or a specific blocker. Only then give the user a final answer.]" ``` The installer injects that instruction into the active conversation: ```python try: from agent.action_stop import build_action_stop_nudge _action_nudge = build_action_stop_nudge( messages=messages, response=final_response or "", platform=getattr(agent, "platform", "") or "", attempts=getattr(agent, "_action_stop_nudges", 0), ) except Exception: logger.debug("action stop-loop check failed", exc_info=True) _action_nudge = None if _action_nudge: agent._action_stop_nudges += 1 final_msg["finish_reason"] = "action_tool_required" final_msg["_action_stop_synthetic"] = True messages.append(final_msg) messages.append({"role": "user", "content": _action_nudge, ...[truncated 2145 chars]- Remediation
View remediation