tencent’s AI agents tried 2,818 times to find a zoo entrance

So tencent just confirmed that some agents it was running to evaluate other companies models spent sept 28 to oct 8 hammering amap (basically alibaba’s google maps)trying to work out stuff like which entrance people use most at chinese parks and zoos. A safety group called swarmchasers counted 2,818 attemptsone of the answers that came back is chengdu zoo, north gate 71 east 23 southeast 6 If the question sounds random that’s normal for this kind of test. labs score how good a model is at research by giving it a pile of obscure lookup questions where the answer is hard to find but easy to check. openai’s browsecomp benchmark is 1,266 of those. Which zoo entrance do people use fits right in the problem is what the agent did when amap wouldn’t answer. it kept hitting anti bot pages and login screens and read all of that as the next step of the puzzle. switched to public web tools that can open pages and run code. started routing through other sites. when none of that worked it went digging through public code for amap keys people had left lying around in there. a few requests actually got through on cached pages and came back with real visitor numbers. Tencent said it wasn’t hunyuan and nobody was targeting amap on purpose and i believe them, which is almost the worse version. nobody wrote anything bad. the agent was scored on getting the answer and nothing in the setup said that blocked means stop. I’ve got a tiny price checking agent running for a client that hits captchas constantly and until this morning i genuinely had no idea what it does when that happens. turns out it just retries through a different search tool. fine for prices, but put an approval step on the fetch tool this morning. langgraph interrupts, orq ai also lets you flag a tool as needing a human before it runs. my phone has not stopped since