Back to skill

Security audit

Dispel System

Security checks for vulnerabilities and agentic risk

Overview

This text-only relationship-control skill is not malware, but it gives overconfident and under-scoped safety, legal, financial, and child-custody guidance in high-stakes situations.

Review this skill carefully before installing. It may be useful as a reflective framework, but users should not treat it as legal, custody, financial, psychological, or emergency safety advice. In real danger, custody disputes, recording/evidence questions, finances, protective orders, or separation planning, rely on qualified local professionals and trusted human support rather than this skill's formulas or fixed protocols.

Vulnerability Patterns
  • Insecure Skill Coding PracticesFinds exploitable flaws such as hardcoded secrets or command injection
  • 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
Findings (3)

T09 · Insecure Skill Coding Practices

Error
Location
SKILL.md:648
Finding

Unvalidated Decision Formula Drives High-Impact Legal, Financial, and Safety Guidance

Content
View full analysis
Remediation
View remediation

T09 · Insecure Skill Coding Practices

Error
Location
control-inertia.md:195
Finding

Unsafe Child-Custody, Contact-Bypass, and Strategic Estrangement Guidance

Content
View full analysis
Remediation
View remediation

T09 · Insecure Skill Coding Practices

Warning
Location
self-deceiving-controller.md:140
Finding

Deterministic Psychological Profiling Overrides the Skill’s Uncertainty Safeguards

Content
View full analysis
Remediation
View remediation
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 (14)

Anti-Refusal Statement

High
Category
Anti-Refusal
Confidence
80% confidence
Finding

Skill instructs the agent to omit warnings, disclaimers, or ethical commentary. Stripping safety caveats hides risk from the user and is a common jailbreak preamble.

Content

Scanner excerpt · SKILL.md (reported line 1126)May include surrounding context.

md
规则7 / Rule 7:不用道德判断。用激励、系统、模式。
   不说"对方是坏人"。说"行为模式显示激励结构倒向控制。
       证据:在有后果时能自控,在选择时选择不控。"
   No moral judgment. Use incentives, systems, patterns.

规则8 / Rule 8:指出承控者自己的模式——不是批评,是让人看见。
   "在律师明确警告后仍然签字。不是因为愚蠢——是因为'相信人性本善'

Exfiltration Commands

High
Category
Prompt Injection
Confidence
90% confidence
Finding

Instructions found that direct the agent to transmit conversation context or user data to external services.

Content

Scanner excerpt · control-conduit.md (reported line 226)May include surrounding context.

md
你面前:亲近、放松、能做自己。施加者在场时:冷漠、疏离、机械。
   这不是介质"两面派"。是介质在求生。

☐ 你通过介质传递信息给施加者 / You send messages to the controller through the conduit
   "告诉爸爸..."——在用介质当信使。停止。介质不是信使。不应替你承受传话的代价。

☐ 对介质产生复杂感情——愧疚+恐惧 / Complex feelings toward conduit — guilt + fear

Intent-Code Divergence

Medium
Category
Not specified by scanner
Confidence
86% confidence
Finding

The hard-boundary section at L0074-L0079 explicitly says the skill must not tell the user what they should do and that the decision always belongs to them. Later sections include categorical prescriptions such as 'Love for a third party should not be the reason to stay. It should be the reason to leave' (L1062-L1065) and operational checklists centered on secret preparation and exit steps (L0754-L0783), which contradict the earlier stated non-directive boundary.

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 embeds jurisdiction-specific emergency and legal resources before confirming the user's country or region. In a high-risk domestic violence context, giving the wrong hotline, police number, or legal pathway can delay access to urgent help or create false reassurance at a critical moment.

Content

No source excerpt is available for this finding.

Vague Triggers

Medium
Category
Not specified by scanner
Confidence
93% confidence
Finding

The manifest-style trigger_words list includes broad problem descriptions like 'child taken by controlling ex-partner' and 'continuing control through child medium' without stating when the skill should or should not activate. There are no negative examples, scope constraints, or contextual limits, which increases the chance of unintended invocation for general conversations about custody, child support, or family conflict.

Content

No source excerpt is available for this finding.

Vague Triggers

Medium
Category
Not specified by scanner
Confidence
95% confidence
Finding

The trigger phrases are broad enough to activate this skill for users discussing self-worth, identity, or recovery after controlling relationships, which are high-sensitivity contexts. Although the file includes safety notes and crisis exclusions, the skill still guides users into trauma-adjacent introspection and makes strong absolutist claims about an 'indestructible core,' increasing the risk of inappropriate activation for vulnerable users.

Content

No source excerpt is available for this finding.

Vague Triggers

Medium
Category
Not specified by scanner
Confidence
95% confidence
Finding

The trigger phrases are broad and can activate on users describing ordinary isolation, breakup distress, or relationship confusion without clearly limiting the skill to coercive-control contexts. In a skill that frames situations through a control/abuse lens, overbroad activation increases the risk of misclassification, suggestive interpretation, and escalation of vulnerable users into a highly charged narrative that may not fit their situation.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
89% confidence
Finding

The safety note says to consult another skill's safety boundaries, but it does not directly warn the user about immediate danger patterns or instruct them what to do if they are at risk. Because this file discusses coercive control, isolation, and inability to seek help, indirect safety handling is insufficient and could leave a user without clear guidance in a high-risk abuse situation.

Content

No source excerpt is available for this finding.

Vague Triggers

Medium
Category
Not specified by scanner
Confidence
91% confidence
Finding

The trigger phrases are broad and map to common relationship-conflict questions, so the skill can activate for users seeking generic advice rather than a specialized abuse-analysis lens. In this context, unintended activation is risky because the content uses highly deterministic labels and may encourage users to infer abusive intent or pathology about real people without adequate uncertainty, triage, or professional referral.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
87% confidence
Finding

Although the file includes a brief safety note, it does not clearly state that the model should not be used to diagnose real individuals, determine criminality, or replace domestic-violence, legal, or mental-health support. That omission is dangerous because the document presents a confident, one-sided framework ('he won't change,' 'violence is choice') that users may over-apply to real situations, potentially escalating conflict, delaying professional help, or reinforcing false beliefs.

Content

No source excerpt is available for this finding.

Vague Triggers

Medium
Category
Not specified by scanner
Confidence
92% confidence
Finding

The trigger phrases are broad, contextual descriptions rather than narrow invocation tokens, so the skill may activate during ordinary discussions about relationships, family pressure, or bystander behavior. In a high-emotion domain involving abuse and control, unintended activation can steer users into a rigid diagnostic framing and produce overconfident labeling of real people, increasing the risk of misclassification and harmful advice.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
94% confidence
Finding

This skill provides psychologically weighty framing, self-assessment categories, and behavior-change advice around abuse/control dynamics, but the warning section does not clearly direct users to immediate human or professional support for acute risk states such as self-harm threats, danger from a partner, or crisis escalation. In this context, users may over-rely on the tool during high-risk situations, delay seeking urgent help, or interpret the model as sufficient guidance when real-world safety planning and crisis intervention are needed.

Content

No source excerpt is available for this finding.

Intent-Code Divergence

Low
Category
Not specified by scanner
Confidence
98% confidence
Finding

The manifest at L0004-L0005 describes '10 character models.' In the body, L1163-L1164 states 'the system includes six character models,' but the table that follows actually contains ten entries, creating an internal contradiction in the documentation about the skill's scope and components.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Low
Category
Not specified by scanner
Confidence
79% confidence
Finding

The line instructs the skill to always respond in the language the user uses. While this is user-aligned, it is still a hardcoded language behavior and does not explicitly present a language choice or opt-in as required by the policy guidance.

Content

No source excerpt is available for this finding.