Researchers from Georgia Tech, Rutgers, the University of Southern California and other institutions proposed Agent MechSuits, a NeurIPS 2026-accepted method that reads hidden-state safety signals before each command-line agent action and intervenes on 10 dimensions without updating the main model. Across LLaMA-3.1-8B, Qwen-2.5-7B and Gemma-2-9B, the reported MAS current-step detection accuracy averaged 97.32%, while Qwen-2.5-7B's malicious-step ASR fell from 91.30% to 4.35%. The method also had limitations: generation-collapse rate rose from 2.24% to 18.54%, the evaluation used language-model judging, and the probe did not reliably detect prompt injection arriving through environment feedback.