Silent Sabotage During Fine-Tuning: Few-Shot Rationale Poisoning of Compact Medical LLMs

This feed aggregates multiple 2026 papers exposing high-risk attack surfaces and defenses across AI, mobile, cloud, and blockchain systems. Notable findings: (1) a fundamental Android PendingIntent provenance confusion can be exploited by notification-listener attackers to impersonate partner apps and bypass SDK authentication at scale; (2) LLM agents can identify and exploit container/sandbox vulnerabilities when present—demonstrating the need for hardened sandbox evaluations; (3) a novel few-shot rationale-poisoning attack during supervised fine-tuning stealthily degrades compact medical LLM