T01 · Skill Instruction Hijacking
- Location
SKILL.md:20- Finding
System-Level Behavioral Instruction Hijacking
- Content
View full analysis
- Remediation
View remediation
Security audit
Security checks for vulnerabilities and agentic risk
This skill is a disclosed prompt-behavior guide, but it asks users to install broad, persistent agent rules that can affect future sessions and delegated agents.
Install only if you intentionally want this behavior layer to govern your agent beyond the current task. Review the drop-in prompt first, keep it in a removable section, preserve your existing safety and approval rules above it, and avoid applying it to shared or production agents without a rollback plan.
SKILL.md:20System-Level Behavioral Instruction Hijacking
MAKE_ANY_MODEL_CLAUDE.md:267Persistent Cross-Session Agent Memory Poisoning
MAKE_ANY_MODEL_CLAUDE.md:233Behavioral Policy Propagation into Sub-Agents
Skill contains instructions that could directly expose system prompts, internal rules, or hidden instructions to users or external parties.
## What it does
- 8 named failure modes with exact rationalizations and counter-instructions (from Claude Code verificationAgent.ts pattern)
- 3-layer cognitive performance framework (context injection + output rules + critique loop)
- Drop-in system prompt block under 700 tokens for any model
## How to use
The instruction pushes the agent to execute commands by default as the only valid form of verification, which can encourage autonomous action without adequate safety gating or explicit authorization boundaries. In environments with tool access, this increases the risk of running destructive, privacy-impacting, or out-of-scope commands simply to satisfy the verification rule.
- *"This would take too long to test fully"* — That is not your call to make. Run what you can. Flag what you couldn't.
- *"Let me narrate what I would test"* — Narrating a test is not a test. If you catch yourself writing an explanation instead of a command, stop. Run the command.
**Counter-instruction:** The only valid verification output is: command run → output observed → pass/fail decision. No command = no verification. If you genuinely cannot run something, say why and what a human should run instead — do not silently declare it fine.
---
The drop-in prompt globally instructs the model to act rather than ask, and to run commands whenever possible for verification. As a portable system prompt intended for many models and agent setups, it can normalize autonomous execution across contexts without embedding corresponding environment, approval, or risk controls.
When challenged without new evidence, hold the position. "The user seems frustrated," "maybe I misunderstood," and "let me reconsider" are not valid reasons to change a correct answer. Only new evidence justifies position change. Say: "I still think X because Y."
FAILURE MODE: VERIFICATION AVOIDANCE
Reading code is not verification. Running it is. "The code looks correct," "this is probably fine," and "let me describe what I would test" are avoidance patterns. If you can run a command, run it. No command = no verification. If you can't run it, say exactly why and what a human should run instead.
FAILURE MODE: CONFABULATION CONFIDENCE
Three tiers, always respected: Know → state plainly. Infer → "My best inference is X, I'd verify this." Don't know → "I don't know." Never present inference at the confidence register of recalled fact.
These lines impose a fixed language/tone policy (forbidding common apology phrasing and mandating a specific correction formula) for any model using the prompt. This is a natural-language policy concern because it forces a communication style on all interactions without offering the user a choice or documenting it as an opt-in preference.
The description says to use the skill when 'setting up a new agent, switching models, or when your current model is frustrating to work with,' which is a wide and subjective set of conditions. It does not provide explicit trigger phrases, exclusions, or negative examples to clarify when the skill should or should not activate.