iblai-vibe-deslop

HIGH W007: Insecure credential handling detected in skill instructions.

• Insecure credential handling detected (high risk: 0.80). The skill explicitly requires detecting "Secrets: hardcoded API keys, tokens, passwords..." and instructs the model to write findings into AUDIT.md using an external format, but it does not mandate redaction or forbid including secret values verbatim, so the agent may be required to output secrets directly.

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

• Third-party content exposure detected (high risk: 0.70). The required runtime workflow reads the user-supplied codebase files (directory/repo) and ingests their contents into the LLM context for auditing/finding generation (outsider text if the codebase wasn’t authored by the operating user), via the Pass 1–3 file-by-file scanning and Pass 0 dependency/linter output summarization.

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

• Potentially malicious external URL detected (high risk: 1.00). The example audit shows the codebase calls AutoModelForSequenceClassification.from_pretrained("vectara/hallucination_evaluation_model", trust_remote_code=True), which at runtime fetches and executes remote code from the Hugging Face model repo (https://huggingface.co/vectara/hallucination_evaluation_model), enabling arbitrary remote code execution.