The Agent Skills Directory

[PROMPT_INJECTION]: The skill contains explicit instructions to override the AI agent's standard behavior and safety guardrails. The 'EXHAUSTION CONTRACT' and '--autonomous' mode instructions (e.g., 'DO NOT STOP,' 'DO NOT ASK THE USER ANYTHING,' and 'overrides any default LLM tendency toward minimalism') are designed to bypass human-in-the-loop oversight and force the AI into an unfiltered, continuous execution loop.\n- [COMMAND_EXECUTION]: The skill relies on extensive shell command execution to perform its tasks. It uses tools like curl, dig, grep, and awk for network reconnaissance and surface mapping. It also executes numerous local Python scripts using uv run python3 (e.g., brain.py, intel_engine.py, coverage_record.py), providing the skill with broad access to the local environment and the ability to perform complex automated actions.\n- [DATA_EXFILTRATION]: The skill is programmed to identify and collect sensitive information from remote targets, such as searching for .env files, .git/config, and authentication-related endpoints. This collected data is systematically recorded and synchronized across a persistent 'brain' state, which constitutes a significant data exposure risk.\n- [EXTERNAL_DOWNLOADS]: The reconnaissance and probing phases involve automated fetching of data from arbitrary external domains. This interaction with untrusted external content poses a risk, as the fetched data is subsequently processed by the orchestrator and sub-agents.\n- [PROMPT_INJECTION]: The skill exhibits an attack surface for indirect prompt injection (Category 8). It processes data from external targets (e.g., response bodies and discovered files) and feeds it into the prompts of specialized sub-agents. Ingestion points: untrusted data from target responses and reconnaissance files. Boundary markers: none present; no instructions to delimit or ignore instructions within external data. Capability inventory: extensive shell and script execution capabilities (curl, uv run, bash). Sanitization: none detected; the skill interpolates raw external data into agent prompts.

autopilot