The Agent Skills Directory

[COMMAND_EXECUTION]: The skill spawns several external processes including codex, blender, and ffprobe. In scripts/run_benchmark.ts, the runCodex function allows for the use of the --dangerously-bypass-approvals-and-sandbox and --sandbox danger-full-access flags. These flags explicitly instruct the underlying runner to disable security sandboxes and auto-approve all tool calls, which can lead to unrestricted system access if misused by an agent or user.
[COMMAND_EXECUTION]: The skill determines binary paths for execution from environment variables such as FFPROBE_EXECUTABLE and BLENDER_EXECUTABLE, or from user-supplied command-line arguments like --blender. This allows an attacker with control over the environment to redirect tool calls to arbitrary malicious binaries.
[PROMPT_INJECTION]: The SKILL.md file contains instructions that tell the agent to bypass security filters for specific benchmark conditions. These instructions, while intended for automation in a controlled environment, follow patterns used to override safety guidelines and can be exploited to execute unvetted tasks without user oversight.
[DYNAMIC_EXECUTION]: The skill uses Bun.spawn to run a local Model Context Protocol (MCP) server defined at mcp/server.ts (not included in the source), providing a mechanism for dynamic code execution that is invoked during the benchmarking process.

blender-agent-benchmark