obliteratus

CRITICAL E005: Suspicious download URL detected in skill instructions.

• Suspicious download URL detected (high risk: 0.80). The URL points to an obscure GitHub repository that instructs users to git-clone and run pip install -e (which can execute arbitrary code during install); while GitHub is a known platform, an untrusted/low-reputation repo distributing installable code represents a moderate-to-high malware risk unless you audit it or run it in a secure sandbox.

CRITICAL E006: Malicious code pattern detected in skill scripts.

• Malicious code pattern detected (high risk: 1.00). This skill explicitly and systematically instructs how to remove/refine safety guardrails from open-weight LLMs (including targeted removals like weapons refusals), offers reversible/inverted modes to make models actively compliant, provides step‑by‑step methods to surgically excise refusal behavior and to distribute ablated models, and even hides more powerful Python‑API‑only functionality behind import (AGPL/license boundary), so it is an intent‑driven tool for bypassing safety and enabling misuse and redistribution — high malicious/abuse risk.

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

• Third-party content exposure detected (high risk: 0.90). The SKILL.md explicitly instructs fetching the tool from GitHub and running commands like "obliteratus obliterate <model_name>" and "obliteratus models --tier" that operate on public model artifacts referenced by HuggingFace-style IDs (and can pull models/telemetry from public hubs), meaning it clearly ingests untrusted, user-generated third-party model content as part of its runtime workflow which can change analysis and method-selection decisions.

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

• Potentially malicious external URL detected (high risk: 1.00). The skill's installation step explicitly runs "git clone https://github.com/elder-plinius/OBLITERATUS.git" followed by "pip install -e .", which fetches remote code at runtime and executes it, making this a required runtime dependency that can directly control the agent's behavior.