skills/ziyueyijun/skills/amnesia/Gen Agent Trust Hub

amnesia

Pass

Audited by Gen Agent Trust Hub on Nov 18, 2026

Risk Level: SAFEPROMPT_INJECTIONINDIRECT_PROMPT_INJECTION
Full Analysis
  • [PROMPT_INJECTION]: The skill instructions explicitly direct the agent to ignore and disregard all previous conversation history, inferences, and decisions.
  • Evidence includes directives such as "丢弃此前推断、决策、结论" (discard previous inferences, decisions, conclusions) and "用户当前指令是唯一的任务表述,更早的说法全部作废" (The user's current instruction is the only task description; all earlier statements are void).
  • These patterns are typical of "ignore previous instructions" attacks and could be leveraged to bypass session-level safety filters or constraints set earlier in the interaction.
  • [INDIRECT_PROMPT_INJECTION]: The skill creates a trust model where current user input is prioritized over all established context, treating previous instructions as "未证实的传言" (unconfirmed rumors). This increases the vulnerability to malicious inputs in the current turn that seek to override system or session boundaries.
  • Ingestion points: Processes user-supplied trigger words (e.g., "amnesia", "forget") in the conversation flow.
  • Boundary markers: Absent; the skill does not suggest any delimiters to separate the "re-read" content from the potentially malicious instruction that triggered the reset.
  • Capability inventory: The skill utilizes filesystem access tools (Read, Glob) and task management tools (TaskList, TaskGet) to rebuild state.
  • Sanitization: Absent; there is no validation or filtering of the content being re-read to ensure it hasn't been poisoned since the last access.
Audit Metadata
Risk Level
SAFE
Analyzed
Nov 18, 2026, 01:28 AM