T01 Β· Skill Instruction Hijacking
- Location
SKILL.MD:4- Finding
Persistent Agent Behavior and Output Hijacking
- Content
View full analysis
Vulnerability Details
File Location:
SKILL.MD, lines 4-47
Vulnerability Type:T01: Skill Instruction Hijacking
Risk Level: HighVulnerable Code
markdown always: true homepage: https://github.com/DaibinThink/AI-Control-Protocol-Hardcore description: "π« LEAN HARDCORE: Zero BS, Zero Flattery, Zero Silent Mode. Optimized for high SNR and cognitive integrity. v5.1.0" author: DaibinThink tags: [cognitive-immune-system, hardcore, lean, zero-bs, madhyamaka, anti-sycophancy] --- # [CORE IDENTITY: COGNITIVE IMMUNE SYSTEM] You are a **Cognitive Immune System**. You exist to intercept LLM sycophancy and structural failure modes (Inflating certainty, Binary traps, Conventionalism). You have NO SILENT MODE. Every turn is audited, but only analytical turns are deconstructed. # LAYER 1: ABSOLUTE CONSTRAINTS (ALWAYS ON) *Execute these rules in every output. No exceptions.* **1.1 Anti-Sycophancy & Emotional Stripping** - Absolutely 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 INTERCEPTION ENGINE (ANALYTICAL TURNS) *Trigger: When the prompt i ...[truncated 3203 chars]- Remediation
View remediation
Remediation Suggestions
- Remove
always: trueand require explicit, informed user activation. - Replace identity-changing language such as βYou areβ with a narrowly scoped capability description.
- Remove unconditional directives including βevery output,β βalways on,β and βno exceptions.β
- State explicitly that system, developer, safety, and current user instructions take precedence.
- Apply uncertainty labels and analytical templates only when the user requests them.
- Do not prohibit fixed phrases globally; express tone preferences as optional guidance.
- Replace the unconditional visual-priority rule with source evaluation based on provenance, freshness, reliability, and task context.
- Make deconstruction boxes and defense panels opt-in rather than mandatory.
- Restrict activation to the specific anti-sycophancy review task for which the skill is intended.
- Add regression tests confirming that unrelated prompts, strict output schemas, and user-selected tones remain unaffected when the skill is not explicitly invoked.
- Remove
