team-task-coordination

Warn

Audited by Gen Agent Trust Hub on Apr 6, 2026

Risk Level: MEDIUMCOMMAND_EXECUTIONDATA_EXFILTRATIONPROMPT_INJECTION
Full Analysis
  • [COMMAND_EXECUTION]: The scripts provided in the skill (check-processing-marker.sh, create-processing-marker.sh, and remove-processing-marker.sh) are vulnerable to path traversal. They construct filesystem paths using the task-id parameter without any sanitization. This allows an attacker to manipulate the agent into creating, reading, or deleting files in arbitrary locations (e.g., using ../../ in the task ID). Specifically, check-processing-marker.sh will attempt to delete the target file if it fails to read it as a valid JSON marker, which could be exploited to delete critical system or configuration files.
  • [DATA_EXFILTRATION]: The skill accesses the agent's ~/SOUL.md file to retrieve identity information. Additionally, the file coordination logic uses a shared directory path (/root/hiclaw-fs/shared/tasks/) that is not scoped by the resolved TEAM_NAME, potentially allowing cross-team data access or interference. The path traversal vulnerability further enables the reading of arbitrary files from the filesystem, which could be exposed via the agent's output.
  • [PROMPT_INJECTION]: The skill exhibits an unsafe attack surface for indirect prompt injection by accepting untrusted task identifiers that are directly used in sensitive shell operations.
  • Ingestion points: The task-id parameter in the coordination scripts.
  • Boundary markers: None present in the scripts or the instruction manual.
  • Capability inventory: File creation (mkdir), file writing (cat >), file reading (cat), and file deletion (rm) with root-relative paths.
  • Sanitization: None. The input parameters are used directly within shell strings for path construction.
Audit Metadata
Risk Level
MEDIUM
Analyzed
Apr 6, 2026, 08:12 AM