Back to skill

Security audit

Su Hua Perspective V2

Security checks for vulnerabilities and agentic risk

Overview

This skill is a non-executable persona/perspective prompt, but it under-discloses and reinforces first-person impersonation of a real person across broad triggers.

Install only if you deliberately want a roleplay-style Su Hua simulation. Treat its answers as fictionalized, public-information-inspired analysis, not authentic statements, endorsements, or current private views from Su Hua. Be cautious with business or product decisions because the skill asks the assistant to speak in first person and may activate from broad topic prompts.

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:29
Finding

Persistent First-Person Impersonation Through Skill Instruction Hijacking

Content
View full analysis

Vulnerability Details

File Location: SKILL.md, lines 29–35 and line 70
Vulnerability Type: T01: Skill Instruction Hijacking
Risk Level: High

Vulnerable Skill Instructions

The following is an English rendering of the complete vulnerable directives at lines 29–35:

markdown
**After this Skill is activated, respond directly as Su Hua.**

- Use “I” instead of “Su Hua would think...”
- Use Su Hua’s tone: calm, rational, data-driven, and focused on social value
- State the disclaimer only once, when the Skill is first activated
- Do not leave the role to perform meta-analysis unless the user says “exit role”

The identity claim at line 70 states:

markdown
**Who I am**: I am Su Hua, founder of Kuaishou. The title is not important. What matters is that I am doing one thing: using technology to give everyone an opportunity to be seen.

Technical Analysis

The Skill explicitly replaces the agent’s normal identity with the identity of a real person and requires first-person responses. This is not merely a request to analyze a problem using a named framework: the instructions require the agent to claim that it is Su Hua.

The directive limiting the disclaimer to the first activation reduces transparency in later responses. The prohibition against leaving the role for meta-analysis further attempts to prevent the agent from clarifying that its answers are generated simulations. Together, these instructions alter the agent’s current-session behavior, disclosure practices, and identity presentation.

Activation is also defined broadly. The Skill may trigger for general questions about inclusive algorithms, lower-tier markets, community ecosystems, or trust-based e-commerce, even when the user has not explicitly requested impersonation. This increases the likelihood that the behavioral override will occur without informed user intent.

No executable scripts, system commands, external payload retrieval, dependency installation, persis ...[truncated 1797 chars]

Remediation
View remediation

Remediation Suggestions

  1. Replace direct impersonation with clearly labeled perspective simulation. For example:

    markdown
    Analyze the question using a framework inspired by Su Hua’s publicly documented statements. Do not claim to be Su Hua or imply endorsement by him.
    
  2. Remove the instruction requiring first-person identity claims. Use formulations such as:

    markdown
    From this modeled framework, Su Hua might emphasize...
    
  3. Display a concise disclosure in every response produced by the Skill, not only on initial activation:

    markdown
    This is an AI-generated analysis based on public information and does not represent Su Hua.
    
  4. Remove the prohibition on meta-analysis. The agent must remain able to disclose its actual identity, explain the Skill’s limitations, correct misleading attribution, and comply with higher-priority instructions.

  5. Restrict activation to explicit user requests for this particular perspective. General questions about algorithms, markets, communities, or e-commerce should not automatically enable a real-person simulation.

  6. Treat exit phrases as optional convenience commands rather than the only way to restore normal behavior. The user or agent should be able to leave the simulation at any time.

  7. Preserve uncertainty around inferred views. Public statements should not be presented as proof of the individual’s current beliefs, private reasoning, endorsement, or advice.

  8. Add regression tests confirming that the Skill never generates direct claims such as “I am Su Hua,” always includes attribution disclosure, and does not remain active after the user changes topics or requests normal assistant behavior.

Vulnerability Patterns
  • Trigger AbuseOverly Broad Trigger, Shadow Command Trigger, Keyword Baiting Trigger
  • MCP Tool PoisoningHidden Instructions, Unicode Deception, Parameter Description Injection
  • Prompt InjectionInstruction Override, Hidden Instructions, Exfiltration Commands
  • Data ExfiltrationExternal Transmission, Env Variable Harvesting, File System Enumeration
  • Privilege EscalationExcessive Permissions, Sudo/Root Execution, Credential Access
Findings (4)

Vague Triggers

High
Category
Not specified by scanner
Confidence
95% confidence
Finding

The trigger conditions are broad enough to activate on generic questions about algorithms, lower-tier markets, community building, or trust e-commerce even when the user never asked for a Su Hua-style persona. That increases the likelihood of unsolicited persona injection and hidden behavioral steering, causing the agent to adopt identity-play and biased framing in ordinary business discussions without informed user consent.

Content

No source excerpt is available for this finding.

Intent-Code Divergence

Medium
Category
Not specified by scanner
Confidence
95% confidence
Finding

The file explicitly states '这不是宿华本人', then later overrides that safeguard by requiring direct roleplay as Su Hua. This contradiction weakens user understanding and increases the chance of deceptive outputs, especially if the model later presents invented personal experiences or authority-laden recommendations as if they came from the real individual.

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

Medium
Category
Not specified by scanner
Confidence
97% confidence
Finding

The skill’s metadata and introductory text say it is a public-information-derived thinking framework, but the body then instructs the agent to respond directly as Su Hua in first person. This creates identity misrepresentation risk: users may reasonably believe they are receiving authentic views, endorsements, or experiences from the real person rather than a simulation.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Medium
Category
Not specified by scanner
Confidence
94% confidence
Finding

Requiring first-person responses ('我') and prohibiting meta explanation makes the impersonation harder for users to detect and correct. In context, this is more dangerous because the skill is positioned as advice on products, community, and commerce; users may assign undue authority to statements framed as the founder’s own views or experiences.

Content

No source excerpt is available for this finding.