drive-session
Pass
Audited by Gen Agent Trust Hub on Nov 11, 2026
Risk Level: SAFECOMMAND_EXECUTIONINDIRECT_PROMPT_INJECTIONEXTERNAL_DOWNLOADSPRIVILEGE_ESCALATION
Full Analysis
- [COMMAND_EXECUTION]: The skill makes extensive use of localized Bash scripts to interact with the environment. It executes
herdrfor tab and agent management,gh(GitHub CLI) for issue tracking,gitfor repository state management, andpsfor process synchronization. - [INDIRECT_PROMPT_INJECTION]: The skill's primary function is to read and act upon the output of other AI agents (supervised sessions). This creates an inherent attack surface where a sub-session could output malicious instructions designed to manipulate the supervisor agent's logic, such as tricking it into merging unauthorized code or bypassing escalation protocols.
- Ingestion points: The
scripts/read.shscript retrieves the terminal output of supervised agents and provides it to the supervisor's context.scripts/frontier.shretrieves issue and dependency data from the GitHub API. - Boundary markers: There are no technical delimiters or sanitization steps used to isolate the sub-session output from the supervisor's instructional context.
- Capability inventory: The supervisor has the power to spawn new agents, send simulated keystrokes and text inputs to sessions, interact with GitHub to manage PRs, and monitor system-wide processes.
- Sanitization: The output retrieved from external sessions is processed directly by the LLM without escaping or filtering.
- [EXTERNAL_DOWNLOADS]: The
scripts/frontier.shscript performs network operations to fetch issue metadata and dependency structures from GitHub's official REST API via theghcommand-line tool. - [PRIVILEGE_ESCALATION]: The
scripts/window.shscript executesps -eo pid=,args=to monitor all active processes on the host machine. This is used to implement a "quiet window" synchronization feature but grants the agent information-gathering capabilities over the entire system state.
Audit Metadata