The Agent Skills Directory

PROMPT_INJECTION (HIGH): The file persuasion-principles.md provides a comprehensive guide for bypassing agent rationalization and internal judgment. It explicitly encourages the use of 'Authority' and 'Commitment' principles to create imperative, non-negotiable instructions like 'YOU MUST' and 'No exceptions' to eliminate 'decision fatigue,' which serves as a blueprint for behavioral override.
PROMPT_INJECTION (MEDIUM): AGENTS_MD_TESTING.md includes 'Variant C' instructions that use emphatic framing and guilt-based compliance ('If you didn't use it, you failed') to coerce the agent into specific actions regardless of the actual necessity of the task.
COMMAND_EXECUTION (HIGH): The script render-graphs.js uses child_process.execSync to pass content extracted directly from a markdown file (SKILL.md) into the system's dot (Graphviz) binary. This presents a code execution risk if the input markdown contains malicious graph attributes or exploits vulnerabilities in the host's Graphviz installation.
DATA_EXFILTRATION (LOW): While not explicitly exfiltrating data, render-graphs.js reads arbitrary file content from the filesystem (based on command-line arguments) and writes processed outputs to a diagrams/ directory. This pattern could be leveraged to read and reformat sensitive documentation into visual formats for later capture.

writing-skills