super-agent
Pass
Audited by Gen Agent Trust Hub on Nov 18, 2026
Risk Level: SAFECOMMAND_EXECUTIONDYNAMIC_EXECUTIONINDIRECT_PROMPT_INJECTION
Full Analysis
- [COMMAND_EXECUTION]: The skill executes arbitrary shell commands defined in the
fleet.jsonconfiguration file viabash -c. These commands are used for status notifications and checking provider quotas. - Evidence:
scripts/notify.shexecutesbash -c "$notify_cmd";scripts/quota.shexecutesbash -c "$cmd". - [DYNAMIC_EXECUTION]: The skill uses a pattern where the source code of internal scripts is pasted into agent messages to be executed on remote machines.
- Evidence:
reference/setup.mdandreference/remote.mddescribe sending the full text ofscripts/probe.shandscripts/worktree.shto remote T3 threads for execution viabash -s. - [INDIRECT_PROMPT_INJECTION]: The skill processes untrusted data from GitHub issues and comments, which can influence its orchestration logic (e.g., routing, intake, and status reporting).
- Ingestion points: GitHub API responses processed in
scripts/intake.sh(issue labels/titles) andscripts/reconcile.sh(comment content/PR state). - Boundary markers: Technical delimiters are absent; the skill relies on instruction-level context management as described in
SKILL.md. - Capability inventory: The skill possesses high-privilege capabilities including local and remote shell execution, GitHub API write access, and thread management.
- Sanitization: Text data is structured using
jq, which provides some character escaping, but the skill lacks specific sanitization for natural language instructions embedded in data.
Audit Metadata