The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill performs a git clone and pip install from an external GitHub repository (elder-plinius/OBLITERATUS) that is not part of the trusted vendor list.
[COMMAND_EXECUTION]: The skill executes multiple shell commands to manage environment setup, hardware checks, and the core abliteration process, including the use of subprocess and direct CLI invocation.
[INDIRECT_PROMPT_INJECTION]: The skill exhibits an attack surface for indirect prompt injection by ingesting model names, paths, and YAML configuration files which are then used in command-line arguments.
Ingestion points: Model identifiers and configuration values provided by the user in SKILL.md and through YAML templates.
Boundary markers: No specific delimiters or validation instructions are provided to ensure user-supplied data doesn't contain command injection characters.
Capability inventory: The skill has broad capabilities to execute shell commands, install packages, and write to the filesystem.
Sanitization: There is no documentation of input sanitization or validation for the model names or paths processed by the skill.
[DATA_EXFILTRATION]: The skill includes an opt-in telemetry feature via the --contribute flag which sends anonymized model names, methods, and metrics to an external research dataset.

obliteratus