goal-batch

Pass

Audited by Gen Agent Trust Hub on Sep 5, 2026

Risk Level: SAFECOMMAND_EXECUTIONDYNAMIC_EXECUTIONINDIRECT_PROMPT_INJECTION
Full Analysis
  • Command Execution: The skill utilizes bash to execute scripts that interact with system utilities like tmux for session isolation, git for version control, and process management tools including kill, setsid, and timeout. These capabilities are used to manage and monitor autonomous work lanes.
  • Dynamic Execution: The launch_lane.sh script dynamically generates a shell wrapper within the system's temporary directory (/tmp) to configure process groups and environment variables for each lane. This generated script is subsequently executed to initiate lane operations.
  • System Configuration Access: The orchestration logic reads local configuration files such as ~/.amplifier/settings.yaml and environment variables. This is used to verify that sub-processes (lanes) will have the necessary settings to communicate with LLM providers before they are launched.
  • Indirect Prompt Injection Surface: The skill processes untrusted repository data and logs, posing an indirect prompt injection surface. Data is ingested from DONE.json and transcript.jsonl files within the workspace. The skill uses session-specific ID markers as boundaries and performs regex sanitization on lane names. Capabilities across all scripts include extensive file writing, git operations, and tmux session management.
  • Process Isolation: A custom tmux shim is created in /tmp/gb-tmux-guard/tmux and added to the PATH of lanes to prevent them from interacting with the shared tmux server, demonstrating a security-conscious design for environment isolation.
Audit Metadata
Risk Level
SAFE
Analyzed
Sep 5, 2026, 04:27 AM
Security Audit — agent-trust-hub — goal-batch