parallel-search-evaluation
Pass
Audited by Gen Agent Trust Hub on Oct 22, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and process untrusted external data, including production workloads, code snippets, and public datasets for evaluation purposes. This represents a potential surface for indirect prompt injection where instructions embedded in the evaluated data could influence the agent's behavior.
- Ingestion points: The skill reads external data from user-supplied workloads and public benchmarks as specified in
SKILL.md, as well as JSONL records defined inreferences/eval-contract.md. - Boundary markers: The instructions include a specific boundary marker directing the agent to "Keep gold answers and annotations out of agent prompts, query generation, and tool messages," which serves to isolate sensitive reference data from the active processing context.
- Capability inventory: The skill has the capability to perform network operations through Parallel's Search and Extract APIs and is tasked with delivering a "runnable harness" which involves code generation.
- Sanitization: While the skill emphasizes data isolation (separating gold labels from prompts), it does not explicitly define sanitization or filtering mechanisms for the input dataset content.
Audit Metadata