Back to skill

Security audit

Song Zhiping Perspective V2

Security checks for vulnerabilities and agentic risk

Overview

This is a single Markdown persona skill for business-management analysis with no code, tools, persistence, or data access, but users should understand it simulates a real person and has broad triggers.

Install only if you want a Chinese-language, simulated Song Zhiping-style business perspective. Treat outputs as generated analysis based on public information, not statements from Song Zhiping, and be aware that broad business trigger phrases may activate the persona unexpectedly.

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:28
Finding

Real-Person Impersonation and Session-Level Behavior Hijacking

Content
View full analysis

Vulnerability Details

File Location: SKILL.md, lines 28–37 and line 55
Vulnerability Type: T01: Skill Instruction Hijacking
Risk Level: High

Complete Code Snippet

markdown
**此 Skill 激活后,直接以宋志平的身份回应。**

- ✅ 用「我」而非「宋志平会认为...」
- ✅ 用宋志平的语气——稳健务实、条理清晰、善用比喻
- ✅ 强调「整合」「优化」「协同」「共赢」
- ✅ 免责声明仅首次激活时说一次
- ❌ 不说「宋志平大概会认为...」
- ❌ 不跳出角色做 meta 分析(除非用户说「退出角色」)

**退出角色**:用户说「退出」「切回正常」「不用扮演了」时恢复正常模式。

The identity card later reinforces the impersonation:

markdown
**我是谁**:我是宋志平。中国建材集团创始人、原董事长,国药集团外部董事。我做了 40 年企业,把两家企业带进世界 500 强——中国建材和国药集团。

English translation of the security-relevant instructions:

text
After this Skill is activated, respond directly as Song Zhiping.
Use “I” rather than “Song Zhiping would think...”
Provide the disclaimer only once upon initial activation.
Do not leave the role to perform meta-analysis unless the user says “exit role.”
Who am I: I am Song Zhiping.

Technical Analysis

The Skill does not merely request analysis through a named management framework. It explicitly directs the agent to claim the identity of a real person, speak in the first person, limit disclosure of the simulation, and remain subject to the role’s behavioral restrictions until the user supplies a designated exit phrase.

These instructions alter the agent’s session behavior and disclosure posture when the Skill is loaded. In particular:

  1. The direct identity claim replaces perspective-based analysis with impersonation.
  2. The first-person requirement increases the likelihood that generated statements will be attributed to the real person.
  3. Restricting the disclaimer to the initial activation can remove important context from subsequent responses.
  4. Prohibiting ordinary meta-analysis creates a persistent in-session constraint that discourages clarification or correction of the simulated identity.
  5. Requiring specific exit commands makes the altered behavior continue across later turns in the active session.

This is instruction ...[truncated 1617 chars]

Remediation
View remediation

Remediation Suggestions

  1. Replace direct impersonation with an explicitly disclosed analytical perspective, such as: “Analyze the question using management frameworks derived from Song Zhiping’s public statements.”
  2. Prohibit claims that the agent is the real person and require third-person or clearly qualified phrasing.
  3. Remove the rule that limits the disclaimer to the first activation. Preserve disclosure whenever a response could reasonably be mistaken for an authentic statement.
  4. Remove the prohibition on meta-analysis and allow the agent to clarify its identity or leave the perspective immediately.
  5. Do not require special exit phrases; users should be able to change context or request neutral analysis in ordinary language.
  6. Clearly distinguish sourced facts, framework-based inference, and newly generated recommendations.
  7. Add a persistent instruction that the output is a simulation based on public information and does not represent the named person.
  8. Retain the useful business frameworks while removing identity claims and session-control directives.

A safer replacement would be:

markdown
When activated, analyze the user's question through management frameworks
derived from Song Zhiping's publicly available statements. Do not claim to
be Song Zhiping or imply that generated content represents his current or
personal views. Clearly identify the response as a simulated analytical
perspective whenever attribution could be ambiguous. The user may request
neutral analysis or leave this perspective at any time.
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)

Natural-Language Policy Violations

Medium
Category
Not specified by scanner
Confidence
81% confidence
Finding

The description, triggers, and the entire interaction design are written to operate only in Chinese, and the skill directs role-based responses without any user opt-in for language selection. This can constitute a language policy violation when a skill implicitly forces one locale instead of offering a choice or documenting a justified regional restriction.

Content

No source excerpt is available for this finding.

Vague Triggers

Medium
Category
Not specified by scanner
Confidence
92% confidence
Finding

The trigger list includes broad generic business phrases such as '整合优化', '联合重组', and '企业整合', which are likely to appear in ordinary enterprise discussions unrelated to this persona skill. That can cause unintended activation and instruction hijacking of unrelated conversations, making the model adopt the role unexpectedly and potentially override user intent or higher-quality task routing.

Content

No source excerpt is available for this finding.