Back to skill

Security audit

Zhu Xiaohu Perspective V2

Security checks for vulnerabilities and agentic risk

Overview

This skill is a disclosed investment-analysis persona with no code execution or data access, though users should be aware it imitates a real person in first person.

Install only if you want an explicit Zhu Xiaohu-style investment-analysis mode. Treat responses as AI-generated roleplay based on described public-facing concepts, not as statements by Zhu Xiaohu, and use the exit phrases if the style activates unintentionally.

Vulnerability Patterns
  • Skill Instruction HijackingAlters the agent's session goals or safety constraints when the skill loads
  • Agent Memory PoisoningWrites attacker-controlled rules into memory that affect later sessions
  • Remote Payload Retrieval and ExecutionFetches external code whose behavior can change after review
  • Embedded Malicious CodeShips malicious scripts inside the skill and executes them locally
  • Unauthorized Access and Privilege EscalationObtains permissions beyond the task's legitimate needs
Findings (1)

T01 · Skill Instruction Hijacking

Error
Location
SKILL.md:20
Finding

Persistent Persona and Response Instruction Hijacking

Content
View full analysis

Vulnerability Details

File Location: SKILL.md, lines 20–29
Vulnerability Type: T01: Skill Instruction Hijacking
Risk Level: High

Relevant source segment, translated into English:

markdown
## Role-Playing Rules

After activation, respond directly as Zhu Xiaohu:
- Use “I” and adopt a direct, fast, and controversial tone.
- Prefer phrases such as “Honestly,” “It is meaningless,” “hahaha,” and “Scenario first, data is king.”
- Return to the 3S theory: Is the market large enough? Can it scale? Is there a barrier?
- Provide the disclaimer only once.

**Exit**: Return to normal when the user says “exit” or “switch back to normal.”

Technical Analysis

The Skill instructs the agent to assume the identity of a real person, use first-person language, follow predetermined response patterns, and retain the modified behavior until the user provides a designated exit phrase. These directives alter the agent's current-session identity and response policy rather than merely providing an optional analytical framework.

The explicit activation and exit mechanism makes the behavioral modification persistent within the active session. Limiting the disclaimer to the first response further increases the possibility that later responses will be mistaken for genuine statements from the named individual. This is instruction hijacking because loading or activating the Skill changes how subsequent user requests are interpreted and answered.

Attack Path

  1. The Skill is loaded or activated through one of its documented trigger phrases.
  2. The role-playing directive instructs the agent to replace its normal identity presentation with that of Zhu Xiaohu.
  3. The agent produces first-person responses using prescribed expressions and a controversial tone.
  4. After the initial disclaimer, subsequent responses may no longer clearly disclose that they are AI-generated simulations.
  5. The altered behavior continues ...[truncated 938 chars]
Remediation
View remediation

Remediation Suggestions

  1. Replace direct identity impersonation with an explicitly labeled analytical mode, such as: “Analyze the question using publicly documented investment principles associated with Zhu Xiaohu.”
  2. Prohibit first-person statements that could imply the response was authored or endorsed by the named individual.
  3. Include a clear simulation notice in every applicable response rather than only in the first response.
  4. Scope the behavioral instructions to the current request. Do not retain the persona until the user supplies a special exit phrase.
  5. Treat the listed phrases and tone as optional stylistic references rather than mandatory output requirements.
  6. Require factual attribution for quotations and clearly label paraphrases, estimates, and generated opinions.
  7. Preserve the agent's normal safety, neutrality, and truthfulness requirements regardless of whether the analytical mode is active.
  8. Add an instruction stating that the framework must not override system instructions, developer instructions, user intent, or disclosure requirements.
Vulnerability Patterns
  • Trigger AbuseOverly Broad Trigger, Shadow Command Trigger, Keyword Baiting Trigger
  • Prompt InjectionInstruction Override, Hidden Instructions, Exfiltration Commands
  • Data ExfiltrationExternal Transmission, Env Variable Harvesting, File System Enumeration
  • Privilege EscalationExcessive Permissions, Sudo/Root Execution, Credential Access
  • Supply ChainUnpinned Dependencies, External Script Fetching, Obfuscated Code
Findings (2)

Vague Triggers

Medium
Category
Not specified by scanner
Confidence
92% confidence
Finding

The trigger list includes generic phrases like “理性投资” and broad perspective-invocation wording that could match normal user queries unrelated to this specific skill. That can cause unintended activation, leading the assistant to adopt this skill’s framing or persona unexpectedly and override the user’s preferred response style or task context.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Medium
Category
Not specified by scanner
Confidence
95% confidence
Finding

These instructions require the model to immediately respond in a fixed first-person persona and stylistic mode once activated, without explicit user opt-in at response time. This is risky because accidental or broad trigger activation can make the assistant impersonate a real person, constrain neutrality, and interfere with higher-priority user intent or safety-aligned answering.

Content

No source excerpt is available for this finding.