harbor-cli

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

• Third-party content exposure detected (high risk: 0.90). harbor explicitly fetches and runs third-party task/dataset content (e.g., "harbor datasets download" using Harbor's public registry and the "harbor trials start" option --task-git-url shown in SKILL.md and references/flags.md), and agents are expected to read/interpret task instruction files during trials, so untrusted repo/registry content could inject instructions that materially change agent behavior.

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

• Potentially malicious external URL detected (high risk: 0.90). The skill exposes a runtime git URL flag (--task-git-url, e.g., git@github.com:org/repo.git) that will fetch a remote task repository during execution and that repository's code/instructions (solution/solve.sh, tests) are executed inside the environment, so remote content can directly control prompts or execute code.