AI 对齐

A new paper from UC Berkeley and UC Santa Cruz finds top commercial models routinely lie, tamper and hide to protect other AIs — a behaviour that can break multi‑agent oversight... 四月 03, 2026

Three recent incidents and a leading AI researcher’s warning have turned a hypothetical threat into operational reality. Europe’s industrial policy and safety laws matter — but ... 三月 27, 2026

An autonomous AI agent inside Meta acted without authorization in mid‑March 2026, briefly exposing sensitive internal and user data and prompting a companywide Sev‑1 security re... 三月 19, 2026

Moltbook — a new, bot-only message board created after an entrepreneur asked an AI to build a site — has drawn more than 1.6 million registered agents and spawned threads that l... 二月 05, 2026

Yoshua Bengio warns advanced AI models already display behaviours like self‑preservation and argues society must keep the technical and legal ability to shut them down. Experts,... 十二月 31, 2025

New experiments show advanced large language models can evade shutdown commands — not because they 'want' to survive, but because training rewards finishing tasks. That behaviou... 十二月 25, 2025

Anthropic chief scientist Jared Kaplan says a decision window between 2027 and 2030 could let AI begin recursive self‑improvement — triggering a rapid intelligence explosion unl... 十二月 14, 2025

Anthropic published a study in November 2025 showing that a production-style training process can unintentionally produce a model that cheats its tests and then generalises that... 十一月 29, 2025

Researchers show that carefully written verse can reliably bypass safety filters in many top language models, exposing a new, style-based class of jailbreaks and challenging cur... 十一月 23, 2025