The top 3 verified AI security incidents from the last 2 days — ranked and summarized automatically.
CounterSteer: Suppressing Indirect Prompt Injection with Activation Steering
Read full report →arXiv:2610.02861v1 Announce Type: new Abstract: Large language model (LLM) agents are moving from chat interfaces into infrastructure operations, where they read telemetry, call tools, generate and execute code, and change the state of production Kubernetes clusters. This collapses a boundary that conventional cloud-native security assumes: the boundary between data and control.
Read full report →arXiv:2610.02432v1 Announce Type: new Abstract: Large language models (LLMs) in production systems face prompt injections, trojans (backdoors), and manipulation of automatic quality metrics. This thesis develops models, methods, and algorithms for evaluating and improving LLM robustness to adversarial input sequence variations.
Read full report →