rlm

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

• Third-party content exposure detected (medium risk: 0.65). The required workflow can stage a repository into the Daytona sandbox via DaytonaInterpreter(repo_url=..., repo_ref=..., context_paths=...), and that repo content (outsider-authored when the URL/ref isn’t the user’s own) becomes readable text inside interp.execute() code, which the LLM then uses—so this is an outsider free-text ingestion path via runtime-fetched repo content.

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

• Potentially malicious external URL detected (high risk: 0.90). The skill passes the runtime parameter repo_url="https://github.com/your-org/your-repo" to DaytonaInterpreter (SKILL.md:134), which causes the sandbox to fetch and stage external repository content at runtime that the interpreter may execute, so the external URL can directly control prompts or executed code.