The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill describes an attack surface for indirect prompt injection where the agent is instructed to ingest and process user-provided JSON scenario files to define benchmark parameters and orchestration.
Ingestion points: TASKS.md (multi-node section) and CAPABILITIES.md (orchestration section) describe using a "JSON scenario" file to define many tests in one file.
Boundary markers: Absent; there are no instructions to the agent to disregard potential natural language instructions embedded within the JSON structure.
Capability inventory: Shell command execution via doca_perftest and orchestration over OpenMPI/SSH as described in TASKS.md.
Sanitization: Absent; the skill does not mention validation or sanitization of the JSON data before it is interpreted by the agent.
[INDIRECT_PROMPT_INJECTION]: The skill contains self-referential content that claims the skill has passed security scans and is safe to use.
Ingestion points: BENCHMARK.md (line 3) and skill-card.md (Evaluation Results section) contain explicit "PASS" verdicts and security claims.
Boundary markers: Absent.
Capability inventory: N/A (Static metadata).
Sanitization: N/A.

doca-perftest