sparring

Pass

Audited by Gen Agent Trust Hub on Nov 5, 2026

Risk Level: SAFECOMMAND_EXECUTIONINDIRECT_PROMPT_INJECTIONDYNAMIC_EXECUTION
Full Analysis
  • [COMMAND_EXECUTION]: The skill is designed to execute external command-line interfaces (CLIs) such as claude and codex to facilitate model-vs-model discussions.
  • The bin/sparctl script acts as a controller that invokes these tools based on the user's request.
  • The lib/tmux-agent.sh script uses tmux to spawn and manage interactive terminal sessions for the agents.
  • The lib/print-agent.sh script executes CLI tools using the timeout command to manage non-interactive turns.
  • [INDIRECT_PROMPT_INJECTION]: The skill possesses an attack surface for indirect prompt injection because it ingests user-provided text and data to generate prompts for an 'opponent' AI model.
  • Ingestion points: User tasks and conversation history are processed in SKILL.md and passed to the sparring harness.
  • Boundary markers: While the skill uses a structured prompt protocol, it lacks explicit safety delimiters to prevent malicious instructions within the user's data from overriding the sparring protocol instructions.
  • Capability inventory: The toolset includes the ability to execute shell commands, manage terminal sessions, and perform file read/write operations in the /tmp/ directory.
  • Sanitization: The scripts perform basic whitespace normalization for terminal input but do not sanitize the semantic content of the prompts to prevent injection attacks against the secondary model.
  • [DYNAMIC_EXECUTION]: The skill assembles and executes shell commands at runtime to interact with the chosen AI backends.
  • In bin/sparctl, the ask-once command constructs a command string by appending a shell-quoted prompt to an agent command and passes it to agent_spawn.
  • The agent_spawn function in lib/tmux-agent.sh executes these dynamically constructed command strings within a new tmux session.
Audit Metadata
Risk Level
SAFE
Analyzed
Nov 5, 2026, 10:52 PM