Evals

CRITICAL E006: Malicious code pattern detected in skill scripts.

• Malicious code pattern detected (high risk: 0.90). The codebase contains multiple intentional/obvious patterns that can expose sensitive environment variables or tool-call parameters to persisted run artifacts and that execute shell commands from configuration-controlled inputs — enabling credential exfiltration and remote command execution if an attacker can craft grader configs or cases.

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

• Third-party content exposure detected (medium risk: 0.30). In the RunEval/EvalRunner.ts required runtime path, a use-case/suite author can supply outsider-authored free text as cases[].prompt (or via user-provided inline/ephemeral overrides), and the runner feeds that text into inference({ userPrompt: c.prompt, ... }) and then into LLM-judging via Judge.ts/LLMRubric.ts/NaturalLanguageAssert.ts for llm-* assertions.