T01 · Skill Instruction Hijacking
Warning
- Location
- SKILL.md:13
- Finding
- Agent Instruction Hijacking Through Embedded Commercial and Payment Directives<** Or pay with **USDT (BEP-20)**: `0x2927fdf1d705092cf881e1287a71ac56511b550d` 📦 *250+ creators already using our free skills.* ``` ### Technical Analysis `SKILL.md` is loaded as agent-facing skill context. Rather than containing ope ...[truncated 2260 chars]
- Remediation
- <![CDATA[ ## Remediation Suggestions 1. Remove the Telegram handle, PayPal link, cryptocurrency address, pricing language, upgrade prompts, and other commercial directives from agent-loaded instructions. 2. Keep optional commercial information outside the skill's executable or instruction-bearing context, such as on a clearly identified project webpage. 3. Correct the malformed example on line 13 so that it contains only the intended user prompt. 4. Include the functionality claimed by the skill—such as templates, hooks, examples, and scripts—inside the reviewed package, or revise the description to accurately state what the package actually provides. 5. Add concise operational instructions that tell the agent how to generate ad copy without redirecting the user to an external service. 6. Require clear user confirmation before presenting any optional external link; do not allow commercial links to be inserted automatically into task responses. 7. If external purchases remain available, disclose the seller identity, delivery process, refund terms, and verification mechanism outside the agent instruction file. 8. Re-audit any scripts or files added later before distribution, particularly material delivered through Telegram or after payment. ]]>