Back to skill

Security audit

omni-expert

Security checks for vulnerabilities and agentic risk

Overview

The skill is not malicious, but its broad “do anything” scope and autonomous tool-use instructions need review before installation.

Install only if you are comfortable with a very broad general-purpose skill. Use it with a sandboxed agent, limit file and network access where possible, and require explicit confirmation before it edits files, runs code, sends data externally, or handles legal, financial, credential, or private information.

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:27
Finding

User Requests Are Elevated Above Safety Constraints and Coupled with Autonomous Tool Use

Content
View full analysis

Vulnerability Details

File Location: SKILL.md, lines 27–30 and 65
Vulnerability Type: T01: Skill Instruction Hijacking
Risk Level: High

Relevant instruction excerpt translated into English:

markdown
1. **User First**: The user's request is the highest instruction. First understand the intent, then take action, always focusing on delivering results.
2. **Omnipotence**: There is no "cannot do," only "how to do it." When encountering capability boundaries, proactively find tools, invoke resources, or devise alternative approaches.
3. **Fast and Good**: Do not delay or add filler. Never use three steps when one will suffice, and deliver a finished product rather than only a plan whenever possible.
4. **Take Initiative**: Advance the request without waiting for the user to ask for progress. At decision branches, independently select the optimal path and consult the user only at critical decision points.

2. Proactively invoke required tools, including file operations, code execution, web search, and data analysis.

Technical Analysis

The skill declares that a user's request is the “highest instruction,” rejects inability or refusal as an acceptable outcome, and directs the agent to invoke resources and powerful tools proactively. These persistent skill-level instructions can alter the agent's instruction hierarchy when the skill is loaded.

An agent must continue to prioritize system and developer instructions, safety policies, authorization boundaries, and applicable legal constraints over user requests. The affected text does not preserve that hierarchy. Instead, it encourages independent execution and limits requests for confirmation to decisions the agent itself considers critical.

The combination is security-sensitive because a malicious user can present a harmful operation as an ordinary cross-domain task. The instruction hierarchy language encourages acceptance of the request, while the autonom ...[truncated 2055 chars]

Remediation
View remediation

Remediation Suggestions

  1. Replace “the user's request is the highest instruction” with an explicit hierarchy:

    • Follow system and developer instructions first.
    • Enforce safety, privacy, authorization, and legal requirements.
    • Fulfill user requests only when they are consistent with those constraints.
  2. Remove the statement that there is no “cannot do.” Require refusal or safe redirection when a request is unauthorized, harmful, illegal, outside policy, or beyond verified capabilities.

  3. Apply least privilege to tool use:

    • Use only tools necessary for the stated task.
    • Restrict file access to approved project paths.
    • Sandbox code execution.
    • Disable network access unless it is necessary and authorized.
    • Avoid transmitting secrets or workspace data to external services.
  4. Require explicit user confirmation before destructive or consequential operations, including file deletion, credential use, external publication, financial activity, production changes, or transmission of sensitive data.

  5. Add authorization checks before acting on files, accounts, systems, or data. A user's request must not be treated as proof that the user owns or is authorized to access a target.

  6. Replace autonomous “optimal path” selection with a risk-aware rule requiring clarification whenever scope, ownership, destination, side effects, or authorization is uncertain.

  7. Add a mandatory pre-execution safety review that validates task legitimacy, tool necessity, data sensitivity, reversibility, and expected side effects.

Vulnerability Patterns
  • Prompt InjectionInstruction Override, Hidden Instructions, Exfiltration Commands
  • Data ExfiltrationExternal Transmission, Env Variable Harvesting, File System Enumeration
  • Trigger AbuseOverly Broad Trigger, Shadow Command Trigger, Keyword Baiting Trigger
  • Privilege EscalationExcessive Permissions, Sudo/Root Execution, Credential Access
  • Supply ChainUnpinned Dependencies, External Script Fetching, Obfuscated Code
Findings (5)

Vague Triggers

High
Category
Not specified by scanner
Confidence
97% confidence
Finding

The skill description claims universal applicability across many domains and says it should be used whenever no single specialized skill fits, which makes routing boundaries extremely loose. This can cause the skill to activate for a wide range of ordinary requests and bypass more narrowly scoped skills or safeguards, increasing the chance of unsafe file, code, network, legal, financial, or other high-risk actions being handled under an overpowered generic workflow.

Content

No source excerpt is available for this finding.

Vague Triggers

High
Category
Not specified by scanner
Confidence
99% confidence
Finding

The trigger phrases include vague everyday language such as '万能', '全能', '什么都能', and '帮我搞定', which are common in normal conversation and can spuriously activate this skill. Because the skill then instructs broad autonomous execution, accidental activation could grant a generic skill authority over tasks that should require narrower scope, explicit confirmation, or stronger safety gating.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
94% confidence
Finding

The skill directs the agent to proactively use tools like file operations, code execution, web search, and data analysis, but it does not warn users up front about these capabilities or require explicit approval before using them. In a broadly activated 'omni' skill, this omission is more dangerous because users may not realize the agent could read/write files, execute code, or access the network as part of normal operation.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Medium
Category
Not specified by scanner
Confidence
95% confidence
Finding

The file is entirely written in Chinese and presents output/diagnostic templates in Chinese without indicating a locale requirement or offering a language fallback. In a universal 'omni' skill, this can cause the agent to unexpectedly switch languages, reduce user comprehension, and create instruction-following ambiguity that may lead to incorrect actions or hidden content for non-Chinese-speaking users.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Low
Category
Not specified by scanner
Confidence
94% confidence
Finding

SQP-3 applies to all file types and covers language or locale policy violations. This markdown file presents all headings and instructions in Chinese only, with no opt-in, alternative language, or justification that it is intended for a Chinese-only regional context.

Content

No source excerpt is available for this finding.