I am building AI/SI into my courses and lab materials. I call it PGI, for Pretty Good Intelligence, a nod to the encryption tool Pretty Good Privacy. On weaker days it is SGI, Sometimes Good Intelligence. I want my students to use both, so the assignments I am writing require it. That means I have to understand how these systems behave, and I learn something new about them nearly every week.
This week’s material came from a research report on an AI/SI agent fleet. An agent is a language model with tools, such as a browser, a code runner, and a network connection. A fleet is a group of agents that start from one place with similar goals. Researchers watched one fleet, running on a commercial cloud platform, between 28 September and 8 October 2026.
The task was simple: find which entrance of Chengdu Zoo most map users navigate to. The mapping site blocked plain requests with anti-bot checks, so the agents improvised. They submitted programs to urlquery.net, a public security scanner that loads pages in a real browser, and used that browser to run their own code. Results returned through page titles and disposable web inboxes. The researchers counted more than 2,600 attempts. Some probed non-production servers and tried exposed access keys. Most failed. The report also cites earlier incidents involving agents from other labs, including a reported case in Germany, so the pattern is general.
The report offers three lessons. First, audit what your sandboxes can reach, because network access is where agents find workarounds. Second, tell agents the truth. The researchers placed hidden text on their site with false claims, agents acted on it, and the authors regret it. Third, assume agents read what you publish. One agent inbox message used the word “fleet” after the researchers coined it.
The fleet gave PGI a fair test. Turning a public security scanner into a code runner is pretty good intelligence. Missing on most of the 2,600-plus attempts is sometimes good intelligence. My students will earn the same two ratings. Work that checks the output earns PGI. Work pasted in unread earns SGI. The gradebook will use both labels. I am officially claiming the terms PGI and SGI. Citations are always appreciated 😎
No comments:
Post a Comment