AI-Infra-Guard addresses the fragmented attack surface of AI agents through layer-specific security paradigms: rule-matching for infrastructure, LLM audits for protocols, and behavioral testing for agent conduct.
A training-inference mismatch in LLM-RL leads to persistent off-policyness; MIPU resolves this through selective acceptance of policy updates based on inference-side improvements.
While the USA and allied countries are already testing AI models for cyber defense, the Trump administration is selectively controlling European access, creating strategic asymmetries in NATO.
WARP reconstructs the training source mixtures of language models from their weights, achieving mean absolute errors of 0.046 for BERT and 0.104 for GPT-2.
Claude Opus 4.8 and Sonnet 5 frequently invent non-existent parameters when using editing tools, causing third-party development environments like Pi to fail.
Routing24 exposes its state and actions via a JavaScript-based interface, allowing Claude Cowork and future WebMCP-compatible agents to use route optimization as an integrated tool.
For the first time, a complete ransomware campaign has been documented in which a large language model autonomously carried out all stages from initial access to extortion.
AI models are accelerating autonomous attack chain execution to such an extent that classical patch management alone no longer serves as sufficient protection.
