rock-agent-sdk

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

• Third-party content exposure detected (high risk: 0.80). This skill explicitly loads datasets from remote/public registries (see the "datasets" section and "RegistryDatasetConfig / RemoteRegistryInfo (default GitHub)" and example "princeton-nlp/SWE-bench_Verified" in SKILL.md), and those third‑party datasets are ingested by Job.run to drive Trials and agent behavior, so untrusted user-generated content could indirectly inject instructions.

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

• Potentially malicious external URL detected (high risk: 0.90). The skill includes an explicit runtime installation command that fetches and executes remote code—"curl https://cursor.com/install | bash"—which will be run during agent setup (sandbox.agent.install), so this URL is a high-confidence risky external dependency.