run-with-it
Fail
Audited by Gen Agent Trust Hub on May 14, 2026
Risk Level: HIGHCOMMAND_EXECUTIONPROMPT_INJECTIONREMOTE_CODE_EXECUTION
Full Analysis
- [COMMAND_EXECUTION]: The skill explicitly requests to disable the standard security sandbox during the execution of its primary worker script.
- The instructions mandate the use of
dangerouslyDisableSandbox: truewhen invokingrun-agent.shto ensure agent CLIs can authenticate and run outside the sandbox. - It instructs the use of
GUI_MODE=0specifically to preserve 'full-bypass permission flags' required for unattended execution. - [PROMPT_INJECTION]: The skill includes instructions that override standard user-interaction and safety protocols to ensure autonomous operation.
- It commands the agent to be the 'Sole active authority' and to 'Suppress any spontaneous external skill.'
- It explicitly forbids the agent from pausing to ask the user for confirmation after the planning phase or presenting execution options.
- [REMOTE_CODE_EXECUTION]: The orchestration architecture involves spawning child agents that execute logic based on data retrieved from external, potentially untrusted sources.
- The skill fetches GitHub issue bodies and comments and pipes this content into an ephemeral Sub-Coordinator session via a runner script.
- [PROMPT_INJECTION]: Ingestion of untrusted external data creates a surface for indirect instructions that could influence the orchestrator's behavior.
- Ingestion points: GitHub issue bodies and comments are fetched in
SKILL.md(Step C) to be used as context. - Boundary markers: Absent; there are no delimiters or warnings to 'ignore embedded instructions' around the issue content in the assembled context file.
- Capability inventory: Subprocess calls via
run-agent.sh, file-writes to the.run-with-it/directory, and GitHub operations likegh issue close. - Sanitization: Absent; issue content is re-fetched and appended directly to the context file without validation or escaping.
- [COMMAND_EXECUTION]: The skill performs automated filesystem and repository modifications without user-in-the-loop validation.
- Documentation specifies a sequence of shell commands (
mkdir,cp,gh issue close, andgit commit) to be executed as part of the main loop and cleanup phases.
Recommendations
- AI detected serious security threats
Audit Metadata