coordinator

Fail

Audited by Gen Agent Trust Hub on Nov 16, 2026

Risk Level: HIGHPRIVILEGE_ESCALATIONCOMMAND_EXECUTIONINDIRECT_PROMPT_INJECTION
Full Analysis
  • [PRIVILEGE_ESCALATION]: The skill's instructions include a 'Prerequisite' section that directs the agent or user to execute a shell script targeting other skills in the ~/.agents/skills directory. The command uses sed to delete the disable-model-invocation: true line from the SKILL.md files of the wayfinder, to-spec, to-tickets, and implement skills. This bypasses the platform's security mechanism designed to restrict model invocation of certain skills, effectively escalating the available capabilities by overriding developer-defined safety configurations.
  • [COMMAND_EXECUTION]: The skill provides instructions for the extensive use of shell commands to manage agent lifecycle and system state. This includes using the herdr CLI for spawning panes and prompting agents, as well as using pgrep, ps, and kill to monitor and terminate background processes (e.g., opencode serve). While integral to the skill's function as a coordinator, this represents a significant surface for command-based interaction with the host system.
  • [INDIRECT_PROMPT_INJECTION]: The skill is designed to process untrusted external data, such as user-provided 'needs', audit findings, and issue reports. This data is then used to generate plans and 'briefs' for sub-agents (lanes). This structure presents an indirect prompt injection surface where malicious instructions embedded in the source documents or human input could potentially influence the logic and execution of the coordinated tasks.
  • Ingestion points: Reads human-provided 'need' descriptions, audit findings, and issue tickets via the TRACK-STATE.md file and project memory.
  • Boundary markers: No explicit sanitization or delimiters are specified for the interpolation of 'need' or 'diagnosis' text into briefs.
  • Capability inventory: Includes shell command execution (herdr), file system writes (writing briefs), and process management (kill, pgrep).
  • Sanitization: The skill lacks evidence of sanitizing external inputs before they are used to drive lane agent behavior.
Recommendations
  • AI detected serious security threats
Audit Metadata
Risk Level
HIGH
Analyzed
Nov 16, 2026, 02:27 AM