Published analysis, newest first. Every entry carries its category, date, and reading time.
In November 2025 Anthropic disclosed GTG-1002, a China-nexus espionage operation that used its own Claude Code agent, wired through MCP tools, to run intrusions against about 30 targets with the AI doing most of the work. What actually happened, how the attackers bypassed the guardrails, and where the reporting outran the evidence.
vLLM, one of the most widely deployed LLM inference servers, shipped a high-severity deserialization bug in its Completions API: crafted prompt embeddings could corrupt server memory even though the code used PyTorch's safe loading mode. What it is, who is exposed, why weights_only=True was not enough, and the fix.
New research from Google's Paradigms of Intelligence team and university collaborators finds that safety training built to stop models claiming consciousness does far more than that: it suppresses mind attribution to animals, nature, and chatbots, and dampens spiritual belief. Ablating one learned direction, or steering a consciousness vector, reverses all of it at once. What the paper shows, how, and what it means for alignment work.
Agentic AI systems make decisions, call tools, and talk to other agents, and the threat models most teams run were never built for any of that. MAESTRO, the Cloud Security Alliance's seven-layer framework for agentic AI, closes that gap. What it is, where classical frameworks fall short, what each layer covers, and how teams are wiring it into CI/CD in 2026.