blender-agent-benchmark

MEDIUM W011: Third-party content exposure detected (indirect prompt injection risk).

• Third-party content exposure detected (medium risk: 0.30). The runtime workflow in scripts/run_benchmark.ts ingests outsider-authored free text from the agent prompt/user task file workdir/TASK.md (generated from task.prompt) into codex exec via runCodex() stdin, and later prompts judges using task.visualBrief/task.visualCriteria (also ultimately free text) into codex exec for judgePair() without selecting a specific pre-approved item from a trusted inbox/feed.