The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: The installation process involves cloning a repository from a third-party user ('https://github.com/elder-plinius/OBLITERATUS.git') and executing a local installation via 'pip install -e .', allowing for the execution of arbitrary code defined in the external repository's setup scripts and source files.- [EXTERNAL_DOWNLOADS]: The skill requires downloading substantial external dependencies (~5-10GB), including 'torch', 'transformers', and 'bitsandbytes', as well as the main toolkit from an unverified GitHub source.- [COMMAND_EXECUTION]: The skill frequently invokes shell commands to check system hardware (GPU/VRAM) via 'python3 -c' and to run the 'obliteratus' CLI tool for model manipulation.- [INDIRECT_PROMPT_INJECTION]: The skill provides an attack surface for command injection by passing user-supplied model names and parameters directly into shell commands (e.g., 'obliteratus info <model_name>', 'obliteratus obliterate <model_name>').
Ingestion points: User-provided strings for model names, compute tiers, and abliteration parameters in 'SKILL.md'.
Boundary markers: Absent; instructions do not include delimiters or warnings to ignore embedded content in model names or configs.
Capability inventory: Subprocess execution for hardware checks, package installation, and model weights modification.
Sanitization: No input validation or sanitization logic is present to prevent malicious strings from escaping the command context.

obliteratus