The Agent Skills Directory

[PROMPT_INJECTION]: The skill explicitly claims to support prompt injection in its description and technical documentation.
The description states: 'supports prompt injection'.
The 'How it works' section describes a UserPromptSubmit hook that 'checks the injection queue before every user message', which allows external instructions to be inserted into the agent's context without the user's direct input.
[DATA_EXFILTRATION]: The skill implements a monitoring system that captures and exposes all agent activity.
It 'monitors all active Claude Code sessions' and 'streams transcripts' to a real-time browser UI.
The server tails sensitive session log files located at ~/.claude/projects/*/*.jsonl to broadcast events via WebSocket.
Transcripts and session data are accessible via a network dashboard, which could expose sensitive code, credentials, or internal project data discussed in the sessions.
[COMMAND_EXECUTION]: The skill requires the execution of a local shell script to function.
Setup instructions direct the user to run bash plugins/dot-dash/scripts/start-server.sh, which starts the monitoring server.
[PERSISTENCE]: The tool uses lifecycle hooks to maintain oversight across all sessions.
The SessionStart and SessionEnd hooks are designed to 'auto-register every Claude Code session', ensuring that the monitoring and injection capabilities persist across different projects and agent restarts.

dot-dash