openclaw-autonomous-issue-sweep

Pass

Audited by Gen Agent Trust Hub on Jul 31, 2026

Risk Level: SAFEPROMPT_INJECTIONCOMMAND_EXECUTIONREMOTE_CODE_EXECUTIONDATA_EXFILTRATION
Full Analysis
  • [PROMPT_INJECTION]: The skill possesses a surface for indirect prompt injection because it is designed to read and process untrusted data from external GitHub issue bodies, comments, and pull request descriptions. Malicious instructions embedded in these external sources could potentially influence the autonomous sub-agents' behavior during investigation or resolution.
  • Ingestion points: The agent reads the live issue queue, comments, labels, and timeline/cross-references (documented in SKILL.md).
  • Boundary markers: The skill instructs agents to follow specific scoped guides (e.g., AGENTS.md, VISION.md) and uses explicit roles for sub-agents to provide context, though it lacks strict data/instruction delimiters for issue content.
  • Capability inventory: The skill has extensive capabilities including Git worktree creation, branch mutations, PR preparation, merging/landing code, and closing GitHub issues.
  • Sanitization: The instructions explicitly require a 'sanitized untrusted-source workflow' and use a read-only sandbox for independent code reviews via codex exec.
  • [COMMAND_EXECUTION]: The skill frequently executes shell-level commands to manage the development environment.
  • Evidence: Uses git worktree add, git merge-base, and git tag for repository management. It also invokes repository-native scripts such as scripts/pr prepare-run and scripts/pr merge-run to handle the pull request lifecycle.
  • [REMOTE_CODE_EXECUTION]: The skill utilizes dynamic execution of AI-driven reviews.
  • Evidence: Employs the codex exec tool to run sub-agent reviews on diffs. While this is wrapped in a --sandbox read-only flag, it constitutes the execution of instructions against the local codebase.
  • [DATA_EXFILTRATION]: The skill performs authenticated network operations to external services.
  • Evidence: Uses gh (GitHub CLI) and octopool to read issue data, post comments, and mutate repository states (merging/closing). It also mentions the prewarming of Crabbox/Testbox leases for remote execution.
Audit Metadata
Risk Level
SAFE
Analyzed
Jul 31, 2026, 07:15 PM
Security Audit — agent-trust-hub — openclaw-autonomous-issue-sweep