agent-long-running-testing
Pass
Audited by Gen Agent Trust Hub on Oct 14, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTIONCOMMAND_EXECUTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill processes complex external data such as agent protocols and execution records, which presents a surface for instructions embedded in analyzed data.
- Ingestion points: Untrusted data enters the agent context through the analysis of Agent, RAG, or LLM plans and evidence sets described in SKILL.md.
- Boundary markers: The skill uses a structured 'ALR-## Finding Contract' and explicit separation between facts and inferences to mitigate accidental obedience to data-embedded instructions.
- Capability inventory: Scripts in evals/fixtures/ perform shell execution and local file access within the evaluation environment.
- Sanitization: The 'Input Audit' protocol in prompts/agent-long-running-testing.md requires the agent to explicitly identify missing or conflicting material.
- [COMMAND_EXECUTION]: The skill includes a shell script evals/fixtures/check-domain-decisions.sh and a Python script evals/fixtures/check_domain_decisions.py intended for deterministic validation of analysis results. These scripts operate on local files and environment variables within the testing scope.
Audit Metadata