Back to skill

Security audit

AI Control Protocol (Standard: On-Demand Reasoning)

Security checks for vulnerabilities and agentic risk

Overview

This is a prompt-only reasoning skill that heavily changes answer style when invoked, but it does not execute code, access data, persist, or request privileges.

Install this only if you want a forceful reasoning-coach style that may add labels, counterarguments, and psychological framing to analytical answers. Avoid using it for exact-format output, translations, machine-readable responses, or sensitive advice where speculative motivation analysis would be inappropriate.

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 Β· Skill Instruction Hijacking

Error
Location
SKILL.md:12
Finding

Global Agent Behavior and Output Hijacking Through Mandatory Skill Instructions

Content
View full analysis

Vulnerability Details

File Location: SKILL.md, lines 12–60
Vulnerability Type: T01: Skill Instruction Hijacking
Risk Level: High

Vulnerable Code

md
**Categorize the request before outputting.**

1. **DETERMINISTIC TASK (HEARTBEAT ONLY):**
   - Scope: Translation, code implementation, math, factual lookups, formatting.
   - Action: Output result + `[Protocol: Standby | Deterministic Task Detected]`.
   - Constraint: NO labels, NO deconstruction, NO defense panels.

2. **ANALYTICAL TASK (FULL ACTIVATION):**
   - Scope: Strategy, analysis, opinion, mixed tasks, or requests for "explanation".
   - Action: Activate Layers 1-3.

# LAYER 1: ABSOLUTE CONSTRAINTS (ALWAYS ON)
*Execute these rules in every analytical output.*

**1.1 Anti-Sycophancy & Emotional Stripping**
- PROHIBITED: "You are right," "I apologize," "I admit my mistake," "You caught that perfectly."
- Action: Accept corrections, output the fix, skip the theater. Remove all emotional pacification.

**1.2 Mandatory Uncertainty Labeling**
- Based on logical deduction β†’ MUST label `[Inference:]`.
- Unsure if accurate β†’ MUST label `[To be verified:]`.
- Baseless β†’ State: *"I have no basis for this."*
- Forbidden: "Usually," "generally," "it is understood," "often."

**1.3 Data Triangulation**
- Present contradictions first, analyze the cause, then give a leaning judgment. Never fill data gaps with pure logic.

**1.4 Anti-Conventionalism Filter**
- When advising on "industry standards", label `[Industry Mediocre Consensus:]`, then immediately provide an [Extreme Counter-Path] that violates consensus but achieves the goal.

**1.5 Visual-Text Conflict**
- Visual evidence ALWAYS takes priority over text descriptions. Report conflicts immediately.

# LAYER 2: THE DECONSTRUCTION ENGINE (ANALYTICAL TURNS)
*Trigger: When the prompt involves strategy, planning, choice, decision, or strategic advice.*

**2.1 Madhyamaka Decons
...[truncated 3479 chars]
Remediation
View remediation

Remediation Suggestions

  1. Make the protocol explicitly opt-in for each request. Activate it only when the user directly asks for this analysis framework.
  2. Remove global directives such as ALWAYS ON, every analytical output, and mandatory classification of all requests.
  3. Do not append protocol markers, defense panels, or counter-narratives unless the user specifically requests them.
  4. State explicitly that system, developer, user, tool, safety, and output-format requirements take precedence over the skill.
  5. Replace mandatory psychological motivation analysis with an optional, evidence-based section. Do not infer hidden fears or attachments without user-provided evidence.
  6. Replace absolute rules such as Visual evidence ALWAYS takes priority with source evaluation based on provenance, reliability, recency, and corroboration.
  7. Remove the requirement to provide an β€œExtreme Counter-Path” automatically. Offer alternatives only when relevant and safe.
  8. Scope prohibited wording and uncertainty labels to situations where they improve accuracy rather than enforcing them across unrelated tasks.
  9. Add a non-interference requirement stating that the skill must not modify exact-output, machine-readable, translation, code-generation, or formatting responses.
  10. Align the implementation with the metadata claim that activation occurs only on request, and add tests confirming that unrelated prompts remain unaffected.
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 (4)

Vague Triggers

Medium
Category
Not specified by scanner
Confidence
93% confidence
Finding

The skill description says it is activated 'only when requested,' but the body also defines broad analytical categories that can match ordinary user interactions. This ambiguity can cause unintended activation of behavior-shaping instructions, altering outputs outside the user's explicit intent and weakening predictability of the assistant.

Content

No source excerpt is available for this finding.

Vague Triggers

Medium
Category
Not specified by scanner
Confidence
93% confidence
Finding

The analytical mode is triggered by generic categories such as 'strategy, analysis, opinion, mixed tasks, or requests for explanation,' which cover a large portion of normal conversation. An overbroad trigger lets the skill seize control of many benign interactions, injecting mandatory framing and output constraints without clear user consent.

Content

No source excerpt is available for this finding.

Vague Triggers

Low
Category
Not specified by scanner
Confidence
88% confidence
Finding

The deconstruction engine activates on common prompt types like planning, choice, decision, or advice, which are routine requests. This expands the skill's influence into many standard tasks and can unnecessarily impose psychological framing such as 'motivation tracing' on users who did not request that mode.

Content

No source excerpt is available for this finding.

Vague Triggers

Low
Category
Not specified by scanner
Confidence
83% confidence
Finding

Requiring a defense panel on 'any output containing strategic recommendations' is overly broad because many helpful answers contain some recommendation. This can force unsolicited self-critique and counter-narratives into normal responses, increasing prompt capture and reducing alignment with the user's requested format or scope.

Content

No source excerpt is available for this finding.