The Agent Skills Directory

[COMMAND_EXECUTION]: The benchmarking scripts (run_benchmark.ts, benchmark_mcp.ts) execute local processes such as the Blender executable and the Codex CLI using Bun.spawn and the Blender Python API. The path to the Blender executable is configurable via CLI arguments, allowing for the execution of arbitrary binaries if misconfigured.
[PRIVILEGE_ESCALATION]: The harness explicitly supports flags that bypass agent security guardrails during benchmarks. Specifically, it uses --dangerously-bypass-approvals-and-sandbox for Codex and --dangerously-skip-permissions for Claude Code to enable non-interactive evaluation. This significantly reduces the security posture of the agent execution environment during testing.
[DATA_EXFILTRATION]: The benchmark isolation logic in scripts/pinned-mcp.ts scans user-specific directories, including the home directory and .codex configuration folders, to identify and programmatically disable other installed skills. This involves broad read access to the local filesystem to ensure benchmark integrity.
[INDIRECT_PROMPT_INJECTION]: The skill implements an automated scoring and comparison pipeline that ingests external data from scene_manifest.json, metrics.json, and agent traces. This creates a surface where a malicious project file or tool output could attempt to influence the benchmark summary or rescoring logic through indirect prompt injection.

blender-agent-benchmark