flow-next-plan-review

Pass

Audited by Gen Agent Trust Hub on May 13, 2026

Risk Level: SAFECOMMAND_EXECUTIONDATA_EXFILTRATIONPROMPT_INJECTION
Full Analysis
  • [COMMAND_EXECUTION]: The skill uses eval to execute shell commands produced by the bundled flowctl script, which is used to initialize session-specific environment variables (W and T) within the agent's shell environment.
  • [DATA_EXFILTRATION]: Project source files and specification documents are transmitted to external analysis services, including RepoPrompt, OpenAI, and GitHub Copilot, for automated review processing.
  • [PROMPT_INJECTION]: The skill's instructions in SKILL.md and workflow.md explicitly command the agent to bypass user confirmation and avoid AskUserQuestion calls during the fix-and-review loop. This behavior, coupled with the ingestion of external reviewer feedback and project code (ingestion points), establishes an indirect prompt injection surface. The skill utilizes boundary markers like <file_contents> and <verdict> tags to delimit content, and its capabilities include executing the flowctl tool for filesystem and network access. Sanitization is performed via regex parsing of the final verdict.
Audit Metadata
Risk Level
SAFE
Analyzed
May 13, 2026, 01:09 AM
Security Audit — agent-trust-hub — flow-next-plan-review