dev-agent-spawn
Fail
Audited by Gen Agent Trust Hub on Mar 15, 2026
Risk Level: HIGHCOMMAND_EXECUTIONPROMPT_INJECTIONDATA_EXFILTRATIONEXTERNAL_DOWNLOADS
Full Analysis
- [PROMPT_INJECTION]: The skill explicitly configures agents to run with dangerous flags such as
--dangerously-skip-permissions,--dangerously-bypass-approvals-and-sandbox, and by disabling sandboxing settings, which bypasses the agent's internal security guardrails. - [COMMAND_EXECUTION]: The main script utilizes AppleScript to automate terminal input, enabling the programmatic execution of arbitrary shell commands in new Ghostty windows, tabs, or splits.
- [COMMAND_EXECUTION]: The
--sendcommand provides a mechanism to inject arbitrary text and keypresses into active terminal sessions viatmux send-keys, allowing for potentially unauthorized command execution. - [DATA_EXFILTRATION]: The
--peekfunctionality usestmux capture-paneto capture and display the content of terminal sessions, which may expose sensitive information, credentials, or private data to the agent or other processes. - [EXTERNAL_DOWNLOADS]: The documentation encourages the use of external, unverified scripts (e.g., in
~/.agent-task-queue/scripts/) that are not part of the skill package, creating a dependency on untrusted local code.
Recommendations
- AI detected serious security threats
Audit Metadata