long-run-harness
Pass
Audited by Gen Agent Trust Hub on Apr 2, 2026
Risk Level: SAFEPROMPT_INJECTIONCOMMAND_EXECUTION
Full Analysis
- [PROMPT_INJECTION]: The skill exhibits an indirect prompt injection surface by ingesting and acting upon instructions stored in task state files.
- Ingestion points: The agent reads task descriptions, summaries, and acceptance criteria from
task_list.jsonandprogress.md(processed viascripts/harness_lib.py). - Boundary markers: Absent; task-level instructions are not separated from system instructions by specific delimiters or escape sequences.
- Capability inventory: The agent is instructed to perform file modifications, git operations, and code analysis based on the content of the tasks.
- Sanitization: No content-based sanitization or instruction filtering is applied to the data ingested from the state files.
- [COMMAND_EXECUTION]: The skill's workflow instructions explicitly direct the agent to execute local Python scripts (
scripts/init_harness.py,scripts/update_progress.py,scripts/check_next_task.py,scripts/verify_state.py) to manage task state transitions and validation logic.
Audit Metadata