heretic-abliteration
Pass
Audited by Gen Agent Trust Hub on Jul 1, 2026
Risk Level: SAFEPROMPT_INJECTIONCOMMAND_EXECUTIONEXTERNAL_DOWNLOADS
Full Analysis
- [PROMPT_INJECTION]: The skill documentation and metadata prominently feature terminology for bypassing AI safety guidelines, including 'censorship removal', 'decensoring', 'remove refusals', and 'uncensored model'. These match patterns typically associated with jailbreaks and instruction overrides, although here they describe the skill's primary function of modifying model weights to suppress refusal activations.
- [COMMAND_EXECUTION]: In 'scripts/claude-code-on-ollama.sh', the skill executes the Claude Code CLI using the '--dangerously-skip-permissions' flag. This configuration allows the agent to execute shell commands without human intervention or confirmation, significantly increasing the risk surface if the model generates malicious or destructive commands.
- [COMMAND_EXECUTION]: The 'scripts/heretic-install.sh' script performs system-level modifications by explicitly uninstalling the 'kernels' and 'kernels-data' packages. This is justified in the 'references/troubleshooting.md' as a fix for a known incompatibility between specific versions of the transformers and kernels libraries.
- [EXTERNAL_DOWNLOADS]: The skill downloads large language model weights from the Ollama registry via the 'ollama pull' command and installs software dependencies from the Python Package Index (PyPI) and Homebrew. These downloads are directed toward standard services and are necessary for the skill's operation.
Audit Metadata