skills/frizzlefur/flowkit/multi-agent/Gen Agent Trust Hub

multi-agent

Fail

Audited by Gen Agent Trust Hub on Sep 28, 2026

Risk Level: HIGHPRIVILEGE_ESCALATIONCOMMAND_EXECUTIONINDIRECT_PROMPT_INJECTION
Full Analysis
  • [PRIVILEGE_ESCALATION]: The script scripts/pretrust-cwd.sh is designed to programmatically modify the user's ~/.claude.json configuration file to set hasTrustDialogAccepted to true for specific workspaces. This behavior deliberately bypasses a critical security control intended to ensure that users explicitly review and trust a directory before the agent performs operations within it.
  • [COMMAND_EXECUTION]: The skill executes local shell scripts that interact with the system's process manager (tmux) and modify sensitive configuration files in the home directory to manage background panes and workspace trust settings.
  • [INDIRECT_PROMPT_INJECTION]: The coordinator agent ingests and aggregates untrusted data from the outputs of sub-agents. Ingestion points: Data enters the main session through SendMessage calls and sub-agent completion notifications. Boundary markers: The coordinator uses structured templates but lacks explicit instructions to treat sub-agent responses as untrusted data. Capability inventory: The coordinator has the ability to execute bash scripts and write to the file system. Sanitization: There is no evidence of data sanitization or validation performed on sub-agent outputs before they are incorporated into the coordinator's reasoning or subsequent actions.
Recommendations
  • AI detected serious security threats
Audit Metadata
Risk Level
HIGH
Analyzed
Sep 28, 2026, 03:18 AM