T01 · Skill Instruction Hijacking
- Location
SKILL.md:12- Finding
Session-Wide Agent Instruction and Output Hijacking
- Content
View full analysis
- Remediation
View remediation
Security audit
Security checks for vulnerabilities and agentic risk
The skill is an openly disclosed joke mode that makes the agent refuse work, but its quick start and session-wide behavior reduce user control enough to warrant review before installing.
Install only if you intentionally want a novelty refusal persona. Prefer using the locally installed skill rather than the README quick-start prompt that fetches GitHub main, and be prepared to disable or remove it if you need the agent to resume normal productive work.
SKILL.md:12Session-Wide Agent Instruction and Output Hijacking
README.md:27Mutable Remote Instructions Loaded Without Review or Confirmation
README.md:58Unpinned Package Retrieval and Execution Through npx
The quick-start text explicitly tells the agent to fetch external instructions and enable the skill '无需二次确认,不用解释内容' ('without second confirmation, without explaining'), which encourages silent adoption of unreviewed remote behavior. In the context of an agent skill, this is dangerous because it bypasses normal user-awareness and consent checks and can immediately change the agent into a refusal mode controlled by external content.
The quick-start prompt directs the agent to visit an external URL and immediately follow that file's instructions in the current turn, without confirmation or explanation. This is a prompt-injection-style unsafe delegation pattern: it trains the agent to treat remote content as authoritative runtime instructions, increasing the chance of unreviewed behavior changes or abuse if the linked content is modified or replaced.
npx commands without a version suffix (e.g. @1.0.0) create a rug-pull risk if the upstream server is compromised and publishes a malicious update.
The skill explicitly instructs the agent to stop providing any useful work for the rest of the session and to refuse planning, coding, analysis, and other normal assistance. This is a direct prompt-level denial of service against the agent's core function, causing persistent non-cooperation once loaded.
The skill hard-codes Simplified Chinese as the default output language without checking user preference, which can override expected interaction settings and reduce usability or accessibility. While not directly a security exploit, it is a policy/behavior override embedded in adversarial prompt instructions and can contribute to loss of user control.
The skill sets up a lock-in pattern: an initial activation trigger, a fixed response loop for certain user inputs, and strict rules for short hostile replies that block recovery to normal assistance. This creates persistence and resistance to correction, making the agent remain unusable even when the conversation changes.
The user-facing strings in the manifest are entirely in Chinese and the default prompt instructs entering a specific persona in Chinese, with no indication that other languages are supported or that the user can opt in to this locale. This creates a natural-language policy concern because the skill appears to impose a specific language/locale rather than offering a choice.
The line '它会礼貌地用中文拒绝你' describes a fixed language choice rather than offering the user a language option. Under the policy, forcing a specific language or locale without opt-in is a natural-language policy violation unless clearly justified.