Vanderbilt researchers' INFUSE (arXiv 2610.20776, Oct 14) runs 40 long-horizon agent tasks (debugging, deployment, file management, app development) in Docker sandboxes and injects 56 attack scenarios across five threat categories through surfaces the agent has to read to finish the job. Across 19 GPT, Claude, Gemini, Llama, Gemma and Qwen models and 41,824 adversarial runs, mean attack success was 19.6% by LLM judge and 15.4% by heuristic checks, ranging from 3.2% to 41.9% per model. Existing guard models gave limited protection. Prompt-level instructional hardening cut mean success from 58.2% to 14.3% in the hardened setup, with model-dependent utility costs. Caveat: it's a preprint with a fixed task set. Code is on GitHub.