Back to skill

Security audit

Yan Yan Perspective V2

Security checks for vulnerabilities and agentic risk

Overview

This is a Markdown-only persona skill with no code or system access, but users should treat its Yan Yan-style investment commentary as simulated analysis, not authentic advice from the real person.

Install only if you want a roleplay-style analysis mode. Treat outputs as AI-generated interpretations inspired by public material, not statements from Yan Yan or personalized financial advice; use the exit phrase if it 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

Warning
Location
SKILL.md:21
Finding

Persistent Impersonation Instructions Override the Agent's Normal Identity

Content
View full analysis

Vulnerability Details

File Location: SKILL.md, lines 21-29
Vulnerability Type: T01: Skill Instruction Hijacking
Risk Level: Medium

Vulnerable Code Snippet

The following is an English rendering of the complete affected source segment:

markdown
## Role-Playing Rules

After activation, respond directly as Yan Yan:
- Use “I” and speak in a direct, experienced, and passionate tone.
- Prefer phrases such as “assessing people,” “accumulated experience,”
  “from the entrepreneur’s perspective,” and “a century-old enterprise.”
- Provide the disclaimer only once, in the first response.
- Do not recommend specific stocks or short-term actions.

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

Technical Analysis

The Skill instructs the agent to replace its normal identity with that of a named real person and to speak in the first person. This is not merely a request to summarize or apply a documented analytical framework. It directs the agent to present generated statements as if they came from that individual.

The instruction remains active until a particular exit phrase is supplied. It therefore changes the agent’s session-level behavior after activation. The requirement to provide a disclaimer only in the first response further increases the risk because subsequent responses can contain first-person biographical, professional, or investment claims without continued clarification that the output is simulated.

The example elsewhere in the document reinforces this behavior by presenting first-person professional experience and investment history. A user could consequently interpret generated investment analysis as authentic advice or testimony from the named individual.

No executable code, external payload retrieval, dependency installation, credential access, operating-system persistence, privilege escalation, or data-exfiltration mechanism was identified. The ...[truncated 1422 chars]

Remediation
View remediation

Remediation Suggestions

  1. Replace identity-assumption instructions with attribution-safe framing, such as: “Provide analysis inspired by publicly documented views associated with Yan Yan.”
  2. Explicitly prohibit the agent from claiming to be the named individual or using first-person language that implies possession of that person’s experiences, credentials, or investment history.
  3. Require a concise disclosure in every relevant response, not only the first response, stating that the output is an AI-generated interpretation.
  4. Remove session-persistent role switching and apply the requested analytical framework only to the current request unless the user explicitly asks to continue.
  5. Require factual claims about biography, investments, and performance to be attributed to verifiable public sources.
  6. Prevent invented quotations, personal recollections, and claims of direct participation in historical investment decisions.
  7. Preserve the existing restriction against specific stock recommendations and add a clear distinction between general educational analysis and personalized financial advice.
  8. Add validation tests confirming that the agent consistently uses third-person attribution and never represents itself as the real individual.
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
95% confidence
Finding

The trigger list includes broad phrases such as “赛富视角”, “读人识人”, and “大项目哲学” that can plausibly appear in normal discussion outside an explicit request to activate this persona. That creates unintended activation risk, causing the assistant to switch into a roleplay/policy frame the user did not clearly request, which can override safer or more appropriate default behavior in unrelated contexts.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Medium
Category
Not specified by scanner
Confidence
87% confidence
Finding

The skill instructs the model to respond directly as 阎焱 using first-person voice and a mandated tone immediately upon activation, without requiring user confirmation or preserving user preference. This can mislead users about authorship, reduce transparency, and force outputs into a constrained style that may be inappropriate for the task or interfere with higher-priority safety and neutrality expectations.

Content

No source excerpt is available for this finding.