Remove shell access from two agents that do not need it and reword PyTorch autograd guidance that AgentShield flagged as encoded-payload-like text. AgentShield remains B/75 while findings drop 316->310 and high findings drop 26->21. Local tests passed 2369/2369; full GitHub Actions matrix green.
Add compact prompt-defense baselines to active ECC prompt surfaces and copied CLAUDE examples. AgentShield prompt-defense findings are now zero; local tests passed 2366/2366.