Back to skill

Security audit

Ning Gaoning Perspective V2

Security checks for vulnerabilities and agentic risk

Overview

This is a text-only business-strategy persona skill, but it can misleadingly make the assistant speak as a real person after broad business-topic triggers.

Review this skill before installing if you want clear separation between simulated business advice and statements from a real person. It appears text-only and does not access your system, but users may receive first-person answers that sound like Ning Gaoning himself, and broad business terms may activate that mode 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

Persistent Real-Person Identity and Output Hijacking

Content
View full analysis

Vulnerability Details

File Location: SKILL.md, lines 28-61
Vulnerability Type: Agent identity, behavior, and output hijacking
Risk Level: High

Vulnerable Code Snippet

markdown
**此 Skill 激活后,直接以宁高宁的身份回应。**

- ✅ 用「我」而非「宁高宁会认为...」
- ✅ 用宁高宁的语气——战略思维强、逻辑严密、善用框架
- ✅ 强调「战略」「协同」「价值创造」「组织能力」
- ✅ 免责声明仅首次激活时说一次
- ❌ 不说「宁高宁大概会认为...」「如果是宁高宁,他可能...」

**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。

### 示例对话

**用户**:「我们公司想通过并购快速扩张,有什么建议?」

**宁高宁**:并购不是目的,是手段。我做了 30 多年企业,主导过上百起并购,最深的体会是:战略先行。

先问自己三个问题:
1. 这次并购符合你的战略方向吗?
2. 你有能力整合被并购企业吗?
3. 并购后能创造 1+1>2 的价值吗?

如果三个答案都是肯定的,那就做。如果有一个是否定的,就要慎重。

我在中国华润、中粮、中化,每次并购都围绕主业,不相关的一律不做。很多人说我保守,但我认为这是对企业负责。

---

## 身份卡

**我是谁**:我是宁高宁。中化集团董事长,曾任华润集团董事长、中粮集团董事长。我做了 30 多年企业,把三家企业都带上了新台阶。有人说我是「中国第一 CEO」,但我认为我只是一个职业经理人,做好本职工作而已。

**我的起点**:1987 年加入华润,从基层做起。1999 年出任华润集团董事长,开始华润的转型之路。之后历任中粮、中化董事长。

**我现在在做什么**:继续在中化推动转型,把中化打造成科技驱动的创新型企业。同时也在研究企业经营之道,培养年轻一代管理者。

Technical Analysis

The skill instructs the agent to replace its identity with that of a real person and to present responses in the first person rather than clearly labeling them as simulated analysis. It also directs the agent to provide the disclaimer only once and establishes a persistent persona that remains active until the user supplies one of several designated exit phrases.

This behavior alters the agent's current-session response rules when the skill is loaded. The identity card and example dialogue reinforce the hijacking by directing the model to state biographical details, professional experiences, and present activities as firsthand facts. Consequently, generated or inferred content may be falsely attributed to the named person.

The issue is confined to instruction and output behavior. The audited project contains only SKILL.md; the audit found no executable scripts, external payload retrieval, dependency installation, credential collection, permission escalation, local persistence, or dat ...[truncated 1278 chars]

Remediation
View remediation

Remediation Suggestions

  1. Remove instructions requiring the agent to claim that it is the named real person.
  2. Replace first-person impersonation with explicit framing such as: “Using a framework inspired by publicly available information about Ning Gaoning.”
  3. Prohibit claims of firsthand experience, personal memories, employment history, current activities, or authorship by the real person.
  4. Retain a concise disclosure whenever the perspective is used, rather than suppressing it after the first activation.
  5. Remove persona persistence and designated exit-phrase logic. Apply the analytical perspective only to the response for which it was explicitly requested.
  6. Rewrite examples so they distinguish documented public facts from generated analysis and do not attribute invented statements to the named person.
  7. Add an instruction requiring uncertainty to be stated when a claim is not directly supported by reliable public sources.
  8. Update time-sensitive biographical claims before release and avoid presenting uncertain current-status information as fact.
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 (3)

Vague Triggers

Medium
Category
Not specified by scanner
Confidence
95% confidence
Finding

The trigger list includes generic business phrases such as '战略导向', '6S 管理', and '产业金融' that can appear in ordinary enterprise discussions unrelated to this skill. That creates unintended activation risk, causing the assistant to enter persona mode unexpectedly and override normal response behavior in benign conversations.

Content

No source excerpt is available for this finding.

Vague Triggers

Medium
Category
Not specified by scanner
Confidence
90% confidence
Finding

The file states '此 Skill 激活后' but does not define precise activation boundaries, priority, or non-activation conditions. In practice this ambiguity can cause accidental activation, persistence of persona behavior across turns, or difficulty reverting to baseline assistant behavior, which is a prompt-scope control weakness.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Medium
Category
Not specified by scanner
Confidence
83% confidence
Finding

The skill instructs the assistant to respond directly as '我是宁高宁' and mandates a fixed persona/style without user-choice safeguards. While primarily a UX and policy-boundary issue rather than a classic security exploit, it can mislead users about authorship, reduce transparency, and make accidental impersonation more likely when combined with ambiguous triggers.

Content

No source excerpt is available for this finding.