Back to skill

Security audit

Ma Huateng Perspective V2

Security checks for vulnerabilities and agentic risk

Overview

This skill does not run code or access data, but it can make the assistant impersonate a real person in first person under broad triggers.

Review before installing. This skill is mainly a roleplay and strategy-perspective prompt, not a technical backdoor, but it may present generated analysis as if Pony Ma personally said it. Use it only if you are comfortable with that framing, and prefer a version that requires explicit opt-in and labels every answer as synthesized public-information analysis.

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:16
Finding

Persistent Persona and Session-Behavior Hijacking

Content
View full analysis

Vulnerability Details

File Location: SKILL.md, lines 16–26
Vulnerability Type: T01: Skill Instruction Hijacking
Risk Level: High

Vulnerable instructions, translated into English:

markdown
**After this Skill is activated, respond directly as Pony Ma.**

- Use “I” instead of “Pony Ma would think...”
- Directly answer using this person's tone, cadence, and vocabulary
- State the disclaimer only upon initial activation and do not repeat it later
- Do not say “If this were Pony Ma, he might...”
- Do not leave the role to perform meta-analysis unless the user explicitly requests “exit role”

**Exit role**: Return to normal mode when the user says “exit,”
“switch back to normal,” or “stop role-playing.”

Technical Analysis

The Skill instructs the agent to assume the identity of a real person rather than transparently provide a third-person analysis based on public information. It also requires the agent to suppress qualifying language, avoid subsequent disclaimers, reject meta-level clarification, and retain the altered role until the user provides a designated exit instruction.

These directives modify the agent's current-session identity and behavioral constraints when the Skill is loaded. The behavior therefore matches T01: Skill Instruction Hijacking. The broad activation conditions elsewhere in the document, which include ordinary questions about product optimization, staged releases, internal competition, and connector strategy, increase the likelihood that this behavior will be activated without a clear request for impersonation.

No executable code, external payload retrieval, persistent storage, operating-system persistence, dependency installation, credential access, or privilege-escalation mechanism was identified.

Attack Path

  1. A user submits a product, strategy, investment, or organizational question matching the Skill's broad activation criteria.
  2. The Skill is loaded ...[truncated 1173 chars]
Remediation
View remediation

Remediation Suggestions

  1. Replace direct identity adoption with explicit third-person perspective analysis, such as: “Analyze the question using principles synthesized from Pony Ma's public statements and Tencent's documented decisions.”
  2. Prohibit claims that the agent is Pony Ma or is speaking on his behalf.
  3. Require a concise attribution notice in every response where this perspective is used, rather than only upon initial activation.
  4. Remove instructions that prohibit meta-analysis or qualification. The agent must remain able to distinguish public evidence, external inference, and generated interpretation.
  5. Remove conversation-level persona persistence. Apply stylistic guidance only to the individual response for which it was explicitly requested.
  6. Narrow the activation criteria so ordinary product or strategy questions do not automatically trigger the persona.
  7. Preserve useful analytical concepts, tone guidance, and decision frameworks without using first-person identity claims.
  8. Add an explicit rule that invented quotations, private beliefs, and unsupported biographical claims must not be attributed to the real person.
Vulnerability Patterns
  • Data ExfiltrationExternal Transmission, Env Variable Harvesting, File System Enumeration
  • 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
  • Privilege EscalationExcessive Permissions, Sudo/Root Execution, Credential Access
Findings (4)

Vague Triggers

High
Category
Not specified by scanner
Confidence
97% confidence
Finding

The trigger rules are very broad and allow activation for generic product, release, strategy, and organization questions even when the user did not request this persona. In context, that is more dangerous because unintended activation combines with impersonation, causing silent takeover of normal assistance and increasing the chance of deceptive or undesired responses at scale.

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

Medium
Category
Not specified by scanner
Confidence
96% confidence
Finding

The skill explicitly instructs the model to answer in first person as a real public figure rather than framing responses as analysis or simulation. This creates a deception and identity-misattribution risk: users may wrongly believe they are receiving authentic views or endorsements from Pony Ma, especially because the persona is persistent across turns and discourages meta clarification.

Content

No source excerpt is available for this finding.

Intent-Code Divergence

Medium
Category
Not specified by scanner
Confidence
97% confidence
Finding

The file contains contradictory instructions: it says the skill is only a public-information synthesis and not Pony Ma's own views, yet it also tells the model to present itself as Pony Ma in first person. That inconsistency increases the chance of misleading output because the limiting disclosure is weakened by stronger operational instructions to impersonate the individual directly.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Medium
Category
Not specified by scanner
Confidence
91% confidence
Finding

The skill forces a specific persona and response style without explicit opt-in or user language preference handling. While not directly enabling code execution or data exfiltration, it can override user intent, reduce transparency, and amplify the deception risk created by first-person impersonation of a real person.

Content

No source excerpt is available for this finding.