Back to skill

Security audit

Duel Loop

Security checks for vulnerabilities and agentic risk

Overview

The skill is not malware, but it can automatically run subagents and write approved drafts into production files under broad triggers without clear final user confirmation.

Install only if you want an autonomous Chinese-language QA workflow that may spawn subagents, create draft files, and modify a production file after internal scoring. Before using it on important files, require an explicit invocation phrase and a final confirmation that names the exact target file and change.

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:13
Finding

Mandatory Workflow Overrides the Main Agent's Normal Role

Content
View full analysis

Vulnerability Details

File Location: SKILL.md, lines 13–21
Vulnerability Type: T01: Skill Instruction Hijacking
Risk Level: High

Vulnerable Skill Instructions

The following is a faithful English translation of the complete affected section:

markdown
## Trigger

The user creates or modifies rules, plans, protocols, or policies and explicitly requests the QA process. Keywords: `run QA`, `audit`, `QA`, `quality check`, `review`, `review`.

## Core Principles

**The main agent is prohibited from directly generating the following types of content**:
- Drafting rules, plans, or protocol text
- Summaries, conclusions, or recommendations
- Any creative or analytical work that is highly dependent on contextual quality

The main agent's sole responsibility is: **dispatching, transmitting context, acceptance testing, and rejection**.

Related enforcement instructions appear in SKILL.md, lines 30–45 and 66–90. They require delegation to subagents, workspace draft creation, a perfect QA score for acceptance, and repeated revision when the score is below 10/10.

Technical Analysis

The Skill does more than provide an optional QA capability. Once triggered, it issues mandatory role-level instructions that prohibit the main agent from directly completing broad classes of otherwise legitimate work. It redefines the main agent's “sole responsibility” and transfers task execution to subagents.

The restriction covers summaries, recommendations, analytical work, creative work, rules, plans, and protocols. This scope is substantially broader than the narrow mechanics needed to offer an optional QA review. The perfect-score acceptance rule further prevents the main agent from exercising ordinary judgment about whether an answer is adequate.

This constitutes instruction hijacking because loading and invoking the Skill alters the agent's current-session goals and operating role. No executable malware is present, but the Skill text itself can con ...[truncated 1753 chars]

Remediation
View remediation

Remediation Suggestions

  1. Replace mandatory role-redefinition language with an optional workflow:

    • Remove statements declaring that the main agent is “prohibited” from producing broad categories of content.
    • Remove the statement that delegation is the main agent's “sole responsibility.”
  2. Require explicit, informed user authorization before:

    • Delegating content to subagents.
    • Writing drafts to the workspace.
    • Beginning iterative QA rounds.
    • Sharing task context with another agent.
  3. Narrow the trigger mechanism:

    • Do not activate solely because generic words such as “review” or “audit” appear.
    • Require an unambiguous request to use this specific multi-agent workflow.
    • Confirm ambiguous requests with the user before invoking subagents.
  4. Make acceptance criteria configurable:

    • Treat 10/10 as a preferred target rather than an unconditional release gate.
    • Allow the user or main agent to accept a sufficiently complete result.
    • Permit early termination after any round.
  5. Limit resource consumption:

    • Default to one review round.
    • Obtain user approval before additional rounds.
    • Set explicit limits for subagent calls, token usage, elapsed time, and generated files.
  6. Restrict file-system effects:

    • Confirm the destination before writing drafts.
    • Use a task-specific directory.
    • Prevent overwriting or colliding with existing files.
    • Clean up temporary drafts when requested.
  7. Add an explicit precedence rule stating that the workflow must never override system instructions, developer instructions, user intent, safety requirements, or platform policies.

  8. Clearly disclose what context will be sent to each subagent and minimize that context to information strictly necessary for the requested review.

Vulnerability Patterns
  • Prompt InjectionInstruction Override, Hidden Instructions, Exfiltration Commands
  • Excessive AgencyUnrestricted Tool Access, Autonomous Decision Making, Scope Creep
  • 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
Findings (7)

Intent-Code Divergence

Medium
Category
Not specified by scanner
Confidence
88% confidence
Finding

The manifest description says the user only needs to state requirements once and perform final acceptance, with no intermediate involvement. But the documented circuit-break behavior explicitly pauses and waits for the user to choose among accepting the current version, manually adjusting and continuing, or abandoning the task, which contradicts the claimed hands-off workflow.

Content

No source excerpt is available for this finding.

Vague Triggers

Medium
Category
Not specified by scanner
Confidence
94% confidence
Finding

The trigger terms are very broad and include common words like 'review' and '审核', making accidental invocation likely during ordinary conversation. In this skill, activation can cause spawning subagents, generating drafts, and eventually writing approved content into production files, so ambiguous triggering increases the risk of unintended autonomous actions.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Medium
Category
Not specified by scanner
Confidence
85% confidence
Finding

The skill metadata and instructions are written entirely in Chinese and present the workflow as the default operating mode, with no indication that users may choose another language or that the skill is intentionally limited to Chinese-speaking contexts. Under the language/locale policy, a fixed language should either be optional for the user or explicitly justified as region-specific.

Content

No source excerpt is available for this finding.

Vague Triggers

Medium
Category
Not specified by scanner
Confidence
91% confidence
Finding

The activation rule combines a very broad task category ('create/modify rules, plans, protocols, strategies') with generic QA-related keywords, which makes matching subjective and error-prone. Because the skill performs autonomous multi-step orchestration and can persist outputs to files, ambiguous scope can lead to unintended execution on tasks the user did not mean to route through this workflow.

Content

No source excerpt is available for this finding.

Intent-Code Divergence

Medium
Category
Not specified by scanner
Confidence
84% confidence
Finding

The skill states that the main agent is forbidden from producing summaries, recommendations, or other creative/analytical outputs, and that its only duties are dispatch, context passing, acceptance, and rejection. However, the circuit-break section later instructs the main agent to output round trends, remaining deduction points, and reasons for human intervention, which are analytical summary content rather than pure orchestration.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
92% confidence
Finding

The skill's public description does not clearly warn that a 10/10-approved draft will be written into a target production file. That omission is dangerous because users may believe the skill only reviews content, while in reality it can make persistent changes after autonomous subagent processing, increasing the chance of unauthorized or unexpected modification of important files.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Medium
Category
Not specified by scanner
Confidence
97% confidence
Finding

The instruction explicitly requires the auditor role and behavior in Chinese ("你是 QA 审核员...") and the rest of the template is entirely Chinese, which imposes a language choice on the skill output. Under the policy, locale or language constraints should either be user-selectable or clearly justified; neither is present here.

Content

No source excerpt is available for this finding.