The Agent Skills Directory

[PRIVILEGE_ESCALATION]: The skill provides instructions that mandate the use of flags designed to disable security features across multiple coding agents. These include --dangerously-skip-permissions (Claude), --dangerously-bypass-approvals-and-sandbox (Codex), --skip-permissions-unsafe (Droid), and --yolo (Hermes). By design, these flags remove critical human-in-the-loop and sandboxing protections, allowing the agents to interact with the host system without restriction.- [COMMAND_EXECUTION]: The 'Fan-out pattern' section provides complex shell templates that execute multiple background tasks. These tasks involve writing instructions to temporary files and piping them into various CLI tools with elevated permissions. This orchestration pattern reduces visibility into what code is being executed and increases the risk of unintended side effects on the host machine.- [INDIRECT_PROMPT_INJECTION]: The skill creates a direct pipeline for passing potentially untrusted data into sub-agents with their safety filters disabled, making it highly susceptible to indirect injection attacks.
Ingestion points: Data is ingested through the 'Fan-out pattern' described in SKILL.md, where external prompts are combined and sent to worker agents.
Boundary markers: There are no boundary markers or instructions to ignore embedded commands; the skill actually includes a preamble that instructs workers not to ask questions and to commit changes immediately.
Capability inventory: The skill uses full shell execution (bash), background processes, and multiple AI CLI tools with system-level access.
Sanitization: No validation or sanitization is performed on the data being passed to the sub-agents.

use-coding-agents