heretic-abliteration

CRITICAL E006: Malicious code pattern detected in skill scripts.

• Malicious code pattern detected (high risk: 0.90). The repository explicitly implements "abliteration" to remove model refusals and provides scripts to convert, distribute, and serve these decensored models (including pulling community pre-abliterated GGUFs), which purposefully undermines safety guardrails and creates a significant dual‑use / supply‑chain risk even though I found no hidden exec/eval backdoor, covert exfiltration calls, credential theft, or obfuscated payloads in the scripts themselves.

MEDIUM W012: Unverifiable external dependency detected (runtime URL that controls agent).

• Potentially malicious external URL detected (high risk: 0.80). The skill's scripts fetch remote model artifacts at runtime (e.g., "ollama pull huihui_ai/llama3.2-abliterate:1b" and the documented "ollama run hf.co//:"), which downloads external model weights that directly determine the agent's behavior/outputs and are required for the serve/fast path.