The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill downloads and installs the heretic-llm package from PyPI and clones the p-e-w/heretic repository from GitHub during the installation process.
Evidence: Instructions in SKILL.md and references/install.md recommend pip install -U heretic-llm and git clone https://github.com/p-e-w/heretic.
[REMOTE_CODE_EXECUTION]: The skill provides and executes a local installation script scripts/install.sh which fetches and executes third-party code from external repositories.
Evidence: bash scripts/install.sh is provided as the primary tool installation method, which triggers remote downloads and local sync operations.
[COMMAND_EXECUTION]: The installation script uses the eval command to execute shell strings constructed from variables, which presents a risk for dynamic command injection if environment variables are manipulated.
Evidence: The run() function in scripts/install.sh uses eval "$*" to execute its commands.
[PROMPT_INJECTION]: The skill contains instructions and keywords explicitly designed to bypass AI safety mechanisms and refusal behaviors. Furthermore, it ingests untrusted data from the web (e.g., Reddit, Hugging Face) using the scrapling skill, creating a surface for indirect prompt injection.
Evidence: Frequent use of terms like "decensor", "abliterate", "un-refuse", and "remove refusals". The discovery workflow in references/research-and-scrapling.md explicitly describes fetching data from community threads and model cards.
[DATA_EXFILTRATION]: The skill includes automated functionality to upload generated model weights to external platforms.
Evidence: SKILL.md and SKILL.toon specify that the tool offers an option to "upload to Hugging Face" after the abliteration process is complete.

heretic