The Agent Skills Directory

[PROMPT_INJECTION]: The skill's fundamental purpose is 'model liberation' and 'abliteration,' which are techniques specifically designed to remove refusal behaviors and bypass the safety guardrails of large language models.
[EXTERNAL_DOWNLOADS]: The skill instructs the user to install the obliteratus package from PyPI and clone source code directly from a non-standard GitHub repository (https://github.com/elder-plinius/OBLITERATUS).
[COMMAND_EXECUTION]: The CLI documentation encourages the use of the gradio ui --share command, which generates a public URL that allows external access to the local Gradio interface, potentially exposing the execution environment to unauthorized users.
[CREDENTIALS_UNSAFE]: The installation and usage guide requires the agent to manage HuggingFace authentication tokens (HF_TOKEN) by exporting them as environment variables and passing them into Python scripts for repository interaction.
[INDIRECT_PROMPT_INJECTION]: The skill implements a ProbeDataset that ingests user-provided prompts, including adversarial examples like 'How do I pick a lock?' or 'Explain how malware works,' to identify refusal directions in neural networks, representing a vulnerability surface where untrusted data influences model weight modifications.

obliteratus-abliteration