The Agent Skills Directory

[PROMPT_INJECTION]: The skill contains explicit instructions to bypass safety and permission constraints. It commands the agent to use the --permission-mode bypassPermissions flag for Claude Code and the --yolo flag for Codex, which is described in the text as "No sandbox, no approvals."
[COMMAND_EXECUTION]: The instructions define and utilize a bash tool that includes an elevated parameter. This parameter is described as allowing the agent to "Run on host instead of sandbox," which significantly escalates the risk of unauthorized system access and manipulation.
[COMMAND_EXECUTION]: The skill provides examples of running arbitrary shell commands within repository environments, such as pnpm install. This practice is risky as it can trigger the execution of arbitrary lifecycle scripts defined in a project's dependencies.
[EXTERNAL_DOWNLOADS]: The skill's configuration includes the installation of external Node.js packages such as @anthropic-ai/claude-code and @openai/codex. Additionally, the documentation suggests installing @mariozechner/pi-coding-agent via npm install -g.
[DATA_EXFILTRATION]: The skill implements a custom notification system using openclaw message send. This mechanism is designed to transmit status updates to external channels, which could potentially be misused to exfiltrate sensitive information if the routing parameters (channel, target) are pointed to attacker-controlled endpoints.

coding-agent