iterate-agent-harness
Pass
Audited by Gen Agent Trust Hub on Jul 24, 2026
Risk Level: SAFEPROMPT_INJECTIONCOMMAND_EXECUTION
Full Analysis
- [PROMPT_INJECTION]: The skill facilitates a process where the agent analyzes its own transcripts and external logs to derive new instructions and scripts. This ingestion of untrusted data introduces a surface for indirect prompt injection (Category 8), as malicious content embedded in logs could influence the modification of the agent's harness.
- Ingestion points: Agent transcripts, CI logs, and state files like
.cursor/*-state.md(SKILL.md). - Boundary markers: None identified; the instructions do not specify delimiters or warnings to ignore embedded content in logs.
- Capability inventory: The skill has the authority to modify internal rules, skills, and subagent configurations, and executes scripts or test suites (SKILL.md).
- Sanitization: No sanitization or validation of the ingested evidence is described before it is used to iterate on the harness.
- [COMMAND_EXECUTION]: The skill instructions involve the creation of executable regression checks and the execution of the project's test suite (e.g.,
npm test). This is a legitimate development workflow but involves runtime script creation and execution. - [SAFE]: The skill is a meta-development tool designed to help developers and agents improve system robustness. It contains no signs of data exfiltration, obfuscation, or hardcoded credentials.
Audit Metadata