skills/h00mankind/skills/arena/Gen Agent Trust Hub

arena

Warn

Audited by Gen Agent Trust Hub on Nov 14, 2026

Risk Level: MEDIUMCOMMAND_EXECUTIONINDIRECT_PROMPT_INJECTIONDYNAMIC_EXECUTIONDATA_EXFILTRATION
Full Analysis
  • [COMMAND_EXECUTION]: The benchmark mode instructions direct the agent to execute external CLI tools such as codex, gemini, and cursor-agent via bash. These commands incorporate the contents of a task prompt file (prompts/task.md), which contains user-supplied text. This could allow for command injection if the user input contains shell metacharacters and the execution environment does not provide sufficient isolation or sanitization.
  • [INDIRECT_PROMPT_INJECTION]: The skill acts as an orchestrator, passing user-defined tasks "word for word" to subagents. It lacks explicit sanitization, validation, or boundary markers to differentiate between the task data and potential malicious instructions embedded within that data, making it vulnerable to indirect prompt injection attacks against the subagents or the judging agent.
  • Ingestion points: User-supplied task arguments in SKILL.md and modes/benchmark.md.
  • Boundary markers: Absent; instructions mandate using the task prompt "word for word".
  • Capability inventory: bash (external tool execution), Agent tool (subagent spawning), and Write tool (artifact and report generation).
  • Sanitization: None mentioned for the interpolation of user tasks into subagent prompts.
  • [DYNAMIC_EXECUTION]: The skill generates temporary agent configuration files (e.g., .claude/agents/bench-opus-high.md) at runtime to bypass limitations in model effort settings. These files are then used to spawn new subagents, which represents a dynamic modification of the agent's operating environment.
  • [DATA_EXFILTRATION]: The skill accesses and reads local configuration files from the user's home directory (e.g., ${CLAUDE_CONFIG_DIR:-$HOME/.claude}/agent-models.md) to determine model routing and role definitions. While this is part of its functional design for benchmarking, it involves accessing sensitive local configuration state.
Audit Metadata
Risk Level
MEDIUM
Analyzed
Nov 14, 2026, 06:04 PM