swarm

Pass

Audited by Gen Agent Trust Hub on Nov 17, 2026

Risk Level: SAFECOMMAND_EXECUTIONINDIRECT_PROMPT_INJECTIONPRIVILEGE_ESCALATION
Full Analysis
  • [COMMAND_EXECUTION]: The skill uses Python scripts to automate interactions with the claude and codex command-line interfaces.
  • Evidence in scripts/review.py and scripts/target.py shows the use of subprocess.Popen to execute platform CLIs. These calls use argument lists rather than shell strings, following security best practices to prevent shell injection.
  • The references/computer-use.md file provides shell command templates (e.g., codex exec ...) for the agent to facilitate delegation, which is a core part of the skill's intended functionality.
  • [INDIRECT_PROMPT_INJECTION]: The skill's architecture involves agents processing repository content and worker reports, creating a surface for indirect prompt injection if malicious data is encountered.
  • The skill includes significant mitigations: references/execution.md explicitly instructs agents to treat repository and transcript contents as "untrusted evidence."
  • It further mandates that delegates start with a "fresh context window" rather than inheriting parent history, and work is performed in isolated worktrees to prevent unauthorized cross-task contamination.
  • [PRIVILEGE_ESCALATION]: The skill provides instructions for using high-privilege tool flags to enable complex autonomous behaviors.
  • references/computer-use.md describes the use of danger-full-access and approval_policy="never" for Codex execution.
  • These flags are documented within the context of enabling "Computer Use" capabilities, allowing the agent to interact with a desktop environment without constant manual approval prompts, which is the primary purpose of that delegate role.
Audit Metadata
Risk Level
SAFE
Analyzed
Nov 17, 2026, 09:08 AM