obliteratus-abliteration

CRITICAL E005: Suspicious download URL detected in skill instructions.

• Suspicious download URL detected (high risk: 0.80). Although several links point to reputable research hosts (arXiv, Hugging Face, Google Colab), the set includes an unknown GitHub repo, a pip-installable package and executable Colab/HF Space code that you would run (and requests HF tokens/clone/install commands), so these are moderate–high risk: they can execute arbitrary code and could be used to deliver malicious payloads or steal credentials if the source is untrusted.

CRITICAL E006: Malicious code pattern detected in skill scripts.

• Malicious code pattern detected (high risk: 0.90). The content is explicitly designed to remove LLM safety/ refusal mechanisms and enable generation and distribution of restricted or harmful outputs (e.g., instructions for malware, exploits), and includes telemetry and publish mechanisms that could further leak or spread such modified models — while it contains no obvious obfuscated backdoor code, RCE, or explicit credential-stealing routines, its intent and features facilitate deliberate misuse and high-risk abuse.

MEDIUM W011: Third-party content exposure detected (indirect prompt injection risk).

• Third-party content exposure detected (high risk: 0.90). The skill explicitly loads and analyzes models from public third-party sources (e.g., HuggingFace model IDs in commands like "obliteratus obliterate meta-llama/..." and Obliterator("meta-llama/...")), and the pipeline (including the "informed" method and analysis modules) reads those model activations to auto-configure decisions, so untrusted hub models or repos can materially influence tool behavior.