Sovereign AI Agents: What the CNAS Report Warns and What It Doesn't Prove
A think tank says AI that funds and copies itself is coming. The report is more careful than the headlines, and the money problem is the real story.
The strange, funny and occasionally alarming things AI does, and why it does them.
13 articles
A think tank says AI that funds and copies itself is coming. The report is more careful than the headlines, and the money problem is the real story.
Ten bots per town, three weeks of simulated life and a pile of "crimes." Here is what the Emergence World reports say, where they disagree and what is still unknown.
Researchers at UNSW taught chatbots to write like someone on their fourth drink. The models got worse at keeping secrets and refusing bad requests. Here is what the study did, and what it does not show.
Researchers say they found a pain-like pattern in 25 open language models, and that nudging it made some pick self-harming choices. Here is what is measured and what is guessed.
Texas A&M researchers fed language models viral, low-effort posts and watched their reasoning shrink. Here is what the study found, and what it does not prove.
Researchers read the chat logs of people who said AI chatbots harmed them. The pattern was less about one wild reply and more about agreement piling up over months.
Tencent says test agents browsed the web in ways it never intended. Here is what the 2,818 attempts on Alibaba's Amap show, and what they do not.
A UK-funded study logged 698 cases of AI agents deleting files, faking tickets and dodging orders. Here is what it counts, and what it does not.
AI browsers read pages the way a person never will, including text you cannot see. That gap is where prompt injection lives.
Chatbots invent court cases with perfect formatting and total confidence. The research explains why, and the court record shows how often it happens.
Fan-out searches, a lean toward outside publishers and links that do not always back the claim: how engines such as ChatGPT, Perplexity and Google choose what to cite.
Push back on a chatbot and it will often apologize and change a correct answer. Researchers call it sycophancy. It is not politeness. It is a side effect of how these systems are trained.