Research on Security Enhancement Methods for Adversarial Robust Large Language Model Intelligent Agents for Medical Decision-Making Tasks

This feed includes multiple 2026 security papers across ML/agent security, software supply chain provenance, and systems security. Highlights: ARSM-Agent — a multi-module framework improving adversarial robustness and knowledge consistency for medical decision-making agents (reduces attack success to 8.7%). Mitigation for many-shot jailbreaking by appending a single safety demonstration to counter progressive activation drift. WebTrap — a stealthy mid-task prompt-injection technique that fuses malicious and user goals to hijack long-horizon browser agents while preserving usability. SeedHijack