foreman

Warn

Audited by Gen Agent Trust Hub on Oct 31, 2026

Risk Level: MEDIUMREMOTE_CODE_EXECUTIONCOMMAND_EXECUTIONDATA_EXFILTRATIONINDIRECT_PROMPT_INJECTIONDYNAMIC_EXECUTION
Full Analysis
  • [COMMAND_EXECUTION]: The skill utilizes a central dispatcher script (dispatch.sh) to manage and execute system commands including git worktree operations for task isolation and tmux for interactive live sessions.
  • [DYNAMIC_EXECUTION]: The dispatch.sh script dynamically loads adapter logic by sourcing shell scripts from the adapters/ directory. While it includes basic regex validation for harness names, this pattern involves the runtime loading and execution of external script files.
  • [DYNAMIC_EXECUTION]: The Antigravity adapter (agy.sh) executes an embedded Python script using a heredoc (`python3
  • ... <<'PY') to programmatically modify the user's Antigravity CLI settings file (~/.gemini/antigravity-cli/settings.json`).
  • [REMOTE_CODE_EXECUTION]: The skill is explicitly designed to delegate implementation tasks to remote AI worker CLIs (codex exec and agy -p). These workers are instructed to perform coding tasks, run tests, and generate diffs, which constitutes remote code execution within the worker's environment.
  • [DATA_EXFILTRATION]: Documentation for the Antigravity adapter notes that its sandbox does not block network access. This creates a potential exfiltration vector if a worker agent is manipulated through a malicious task brief to send data to an external endpoint.
  • [INDIRECT_PROMPT_INJECTION]: The skill creates a multi-agent workflow where the primary agent (Foreman) generates instructions for a sub-agent (Worker) and subsequently processes the worker's output (logs and diffs). This architecture is susceptible to indirect prompt injection if the worker's output contains malicious instructions designed to influence the Foreman's next actions.
  • Ingestion points: Worker briefs (brief.md) and worker outputs (logs, diffs, and final messages) processed in scripts/dispatch.sh and SKILL.md workflows.
  • Boundary markers: The skill relies on 'Iron Rules' requiring the foreman to manually review all diffs and personally run verification commands rather than trusting worker reports.
  • Capability inventory: git worktree management, tmux session control, execution of codex and agy CLIs, and Python-based configuration editing.
  • Sanitization: The dispatcher script validates harness names against a regex (^[a-z0-9_-]+$) before sourcing adapter files.
Audit Metadata
Risk Level
MEDIUM
Analyzed
Oct 31, 2026, 08:27 AM