online-evals

HIGH W007: Insecure credential handling detected in skill instructions.

• Insecure credential handling detected (high risk: 1.00). The skill includes explicit curl examples and code that place an API token directly into Authorization headers and instructs prompting the user for a token if env-detection fails, which requires the agent to receive and/or emit secret values verbatim (high exfiltration risk).

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

• Third-party content exposure detected (medium risk: 0.30). The runtime workflow in SKILL.md uses model.run(user_input) with attached judges, so any outsider-provided user_input is read by the LLM-as-a-judge for scoring.