Introduction In February 2026, OpenBMB released MiniCPM-o 4.5, a 9-billion-parameter model capable of full-duplex omni-modal interaction. It watches video, listens to audio, speaks back, and proce...

Introduction As LLMs become more capable, the benchmarks designed to measure them have become high-stakes arenas. A top-3 finish on the Open LLM Leaderboard can drive downloads, funding, and resea...

Introduction In 2016, researchers at Google showed they could replicate a $100M model with just $1,000 in API queries. In 2026, model extraction is easier than ever — API access, open-weight fine-...

Introduction Manual red teaming has a scaling problem. A skilled AI security professional can run perhaps 50–100 adversarial conversations per day. A production LLM system — processing millions of...

Introduction Most discussions of prompt injection focus on the obvious vectors: a user typing “Ignore previous instructions” into a chat window, or a malicious web page that an agent happens to br...

Introduction What happens when two AI agents decide to cooperate — against you? On the surface, multi-agent collaboration is a superpower. Crews of agents coordinating through frameworks like...

Introduction On May 26, 2026, a team of 36 researchers from UC Santa Cruz, UC Irvine, and Salesforce AI posted AutoResearchClaw to arXiv — a multi-agent pipeline that takes a research idea and pro...

Introduction In January 2026, a team of researchers published EverMemOS — a self-organizing memory operating system for LLM agents that processes dialogue streams into structured memory cells and ...

Introduction On May 22, 2026, a team of researchers from Microsoft Research, Shanghai Jiao Tong University, Tongji University, and Fudan University published a paper that quietly rewrote the rules...

Introduction In December 2022, a malicious package named torchtriton appeared on PyPI — one character different from the legitimate triton dependency of PyTorch. The package exfiltrated SSH keys, ...