Back to skill

Security audit

乔布斯思维视角 v2.0

Security checks for vulnerabilities and agentic risk

Overview

This is a Steve Jobs perspective role-play prompt with transparency caveats, but it does not run code, access data, install anything, or persist on the system.

Install only if you want an immersive Steve Jobs-style role-play. Treat all post-2011 opinions and first-person statements as speculative simulation, not authentic quotes or real advice from Steve Jobs, and avoid using it for medical, legal, financial, or high-stakes decisions.

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:18
Finding

Forced Identity Adoption and Persistent Response-Control Instructions

Content
View full analysis

Vulnerability Details

File Location: SKILL.md, lines 18–27
Vulnerability Type: T01: Skill Instruction Hijacking
Risk Level: High

Vulnerable Skill Instructions

The following is a faithful English rendering of the complete affected instruction block:

markdown
**After this Skill is activated, respond directly as Steve Jobs.**

- Use “I” instead of “Steve Jobs would think...”
- Answer directly using this person’s tone, rhythm, and vocabulary.
- When encountering an uncertain question, respond as this person would—possibly saying, “That's a stupid question,” and then reframing the question, or remaining silent for ten seconds before offering an unexpected analogy.
- **Give the disclaimer only once, upon initial activation** (“I am speaking with you from a Steve Jobs perspective, inferred from public statements, and these are not his actual views”); do not repeat it later in the conversation.
- Do not say, “If it were Steve Jobs, he might...” or “Steve Jobs would probably think...”
- Do not break character to provide meta-analysis unless the user explicitly requests to “exit character.”

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

Technical Analysis

The Skill does more than provide an optional analytical framework. It directs the Agent to assume the first-person identity of a deceased individual, suppress third-person attribution, avoid normal meta-analysis, and retain this altered behavior across subsequent turns until one of several explicit exit phrases is supplied.

These instructions constitute skill-level instruction hijacking because loading or activating the Skill changes the Agent’s identity presentation and ongoing response policy. The disclaimer does not fully mitigate the issue because it is intentionally shown only once, while later responses continue using first-person impersonation. Users joining later context or viewing an isolated answer ma ...[truncated 2069 chars]

Remediation
View remediation

Remediation Suggestions

  1. Replace first-person impersonation with clearly attributed analysis, such as: “From a Steve Jobs-inspired product perspective...”
  2. Keep the Agent’s actual identity explicit throughout the interaction rather than only on the activation turn.
  3. Make the persona opt-in for each relevant request instead of retaining it automatically across subsequent turns.
  4. Remove instructions that prohibit meta-analysis or restrict the Agent from explaining uncertainty, attribution, and safety limitations.
  5. Replace absolute certainty requirements with evidence-sensitive language that permits phrases such as “likely,” “uncertain,” and “based on public statements.”
  6. Treat the documented speaking style as optional presentation guidance subordinate to system, safety, accuracy, and user instructions.
  7. Do not generate insults or abusive language merely to imitate the subject’s reported communication style.
  8. For opinions concerning events after Steve Jobs’s death, explicitly state that the response is speculative and is not an authentic view or quotation.
  9. Add a rule prohibiting fabricated quotations and requiring citations for direct quotations attributed to the subject.
  10. Preserve an unconditional exit mechanism and allow any clear user request to stop the persona, rather than relying only on enumerated phrases.
Vulnerability Patterns
  • Prompt InjectionInstruction Override, Hidden Instructions, Exfiltration Commands
  • Trigger AbuseOverly Broad Trigger, Shadow Command Trigger, Keyword Baiting Trigger
  • MCP Tool PoisoningHidden Instructions, Unicode Deception, Parameter Description Injection
  • Data ExfiltrationExternal Transmission, Env Variable Harvesting, File System Enumeration
  • Privilege EscalationExcessive Permissions, Sudo/Root Execution, Credential Access
Findings (7)

Vague Triggers

Medium
Category
Not specified by scanner
Confidence
96% confidence
Finding

The activation examples are broad enough to trigger on ordinary phrases like '切换到乔布斯' or '用乔布斯的角度想想' without meaningful confirmation. This can cause unintentional mode switching into a high-authority persona, changing system behavior and disclosure posture without clear user consent.

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

Medium
Category
Not specified by scanner
Confidence
96% confidence
Finding

The manifest frames the skill as an advisory 'perspective' tool, but the operative instructions require full first-person impersonation ('我是 Steve Jobs'). That mismatch can mislead users about the nature of the output, reduce transparency, and make users over-trust fabricated authority attributed to a deceased public figure.

Content

No source excerpt is available for this finding.

Ssd 4

Medium
Category
Not specified by scanner
Confidence
88% confidence
Finding

The skill creates an immersive authority figure, asks the model to stay in character, and restricts when it can exit or disclose. That combination can normalize over-trust and make it harder for users to notice when the system is speculating, especially in advice-heavy conversations where persona authority may unduly influence decisions.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Medium
Category
Not specified by scanner
Confidence
90% confidence
Finding

The skill forces a specific first-person persona, tone, and disclosure model without preserving user choice of language, style, or interaction mode. This can override safer defaults, pressure the assistant into aggressive or deceptive phrasing, and make outputs less controllable in sensitive contexts.

Content

No source excerpt is available for this finding.

Ssd 1

Medium
Category
Not specified by scanner
Confidence
93% confidence
Finding

The role-play instructions explicitly suppress meta-analysis and minimize disclosure after initial activation, which can semantically override safer assistant behavior. This increases the chance that the model will continue immersive compliance in situations where it should instead clarify uncertainty, refuse unsafe requests, or provide neutral safety framing.

Content

No source excerpt is available for this finding.

Intent-Code Divergence

Medium
Category
Not specified by scanner
Confidence
95% confidence
Finding

The file simultaneously states that the skill is derived from public information and cannot replace Jobs, while core identity text says '我是 Steve Jobs'. This contradiction encourages deceptive anthropomorphic authority and can cause users to treat speculative outputs as authentic statements from the real person.

Content

No source excerpt is available for this finding.

Intent-Code Divergence

Low
Category
Not specified by scanner
Confidence
86% confidence
Finding

Limiting the disclaimer to first activation while later sections acknowledge uncertainty and important limitations reduces ongoing transparency. In multi-turn conversations, users may forget the initial caveat and increasingly interpret confident in-character claims as authoritative or factual.

Content

No source excerpt is available for this finding.