long-running-harness
Pass
Audited by Gen Agent Trust Hub on Jun 23, 2026
Risk Level: SAFE
Full Analysis
- [SAFE]: The skill defines an architectural pattern for multi-agent collaboration and does not contain any executable scripts, binary files, or automated installation procedures.
- [SAFE]: Data management is handled through a local directory (
.harness/) for project-specific state. No sensitive system paths, environment variables, or credentials are accessed or exposed. - [SAFE]: External references point to official Anthropic engineering documentation regarding harness design patterns.
- [SAFE]: The instructions for context resets and handoffs use standard file operations and do not involve persistence mechanisms or privilege escalation.
- [INDIRECT_PROMPT_INJECTION]: The skill describes an architecture with an inherent attack surface for indirect prompt injection, as agents read files (e.g.,
.harness/spec.md,.harness/sprint-contract.md) generated by other agents. While the provided templates do not use boundary markers or sanitization, this is a structural characteristic of the multi-agent design pattern and not a malicious implementation.
Audit Metadata