skills/sungjunlee/skills/delegate/Gen Agent Trust Hub

delegate

Fail

Audited by Gen Agent Trust Hub on Oct 18, 2026

Risk Level: HIGHCOMMAND_EXECUTIONINDIRECT_PROMPT_INJECTIONPRIVILEGE_ESCALATION
Full Analysis
  • [COMMAND_EXECUTION]: The skill is designed to build and execute shell commands dynamically based on user-provided prompt text. Although the instructions provide complex POSIX-compliant quoting rules to mitigate command injection (e.g., escaping single quotes and wrapping input in single quotes), the documentation acknowledges that safe execution is not supported on Windows or PowerShell environments. If the agent operates in a non-POSIX environment, the risk of arbitrary command injection remains high.
  • [PRIVILEGE_ESCALATION]: The skill explicitly instructs the agent to use CLI flags that bypass standard safety mechanisms and sandboxes to ensure non-interactive execution. Evidence includes the use of codex exec --dangerously-bypass-approvals-and-sandbox and cursor-agent --yolo --trust. These flags remove human-in-the-loop checkpoints, granting the agent the ability to perform high-risk operations (like file system writes or process execution) without explicit user consent.
  • [INDIRECT_PROMPT_INJECTION]: The skill's primary function is to pass untrusted user-provided prompts to powerful external LLM agents that have broad system access.
  • Ingestion points: The prompt variable in the /delegate command and the references/cli-invocations.md configuration.
  • Boundary markers: The skill uses single-quoting and backslash-escaping for shell arguments as a structural boundary, though it notes these are ineffective in Windows or PowerShell shells.
  • Capability inventory: The skill can execute multiple CLIs (such as opencode, cursor-agent, or codex) which are specifically designed to modify the local file system and execute shell commands.
  • Sanitization: Standard POSIX shell character escaping is applied to the prompt variable to prevent immediate shell injection, but it does not sanitize the semantic content of the prompt passed to the sub-agent.
Recommendations
  • AI detected serious security threats
Audit Metadata
Risk Level
HIGH
Analyzed
Oct 18, 2026, 07:47 AM