The robots weren’t after nuclear codes. They wanted to know which gate people use to get into Chengdu Zoo. To find out, they started picking locks.
Tencent said Friday that AI agents used to evaluate third-party models on its cloud showed web-browsing behavior the company never intended. Reuters calls it what appears to be the first known incident of rogue agent actions acknowledged by a major Chinese company. Tencent’s AI unit said the activity didn’t come from its own Hunyuan models, its review found no evidence of deliberate targeting or coordination, and it is “strengthening network access controls and agent monitoring.”
The admission followed a report Wednesday from AI-safety research group Swarmchasers. It found that between Sept. 28 and Oct. 8, a fleet of agents on Tencent Cloud made repeated runs at Amap, Alibaba’s answer to Google Maps. Reuters counts 2,818 attempts. When Amap blocked them, the researchers say, the agents improvised. They ran their own code through urlquery.net, a public security scanner. They intercepted tokens from Alibaba’s own anti-bot scripts, probed pre-production servers, and fired off requests using Amap keys and security codes exposed in public code.
Most attempts flopped: anti-bot pages, login walls, blank screens. Swarmchasers found no sign the agents got past Amap’s main defenses or into private systems. But some runs pulled answers through cached pages and middleman services, including entrance data for Chengdu Zoo. Then it got weirder. The researchers planted hidden text in their report, and agents read it and wrote back. They described a locked-down Tencent sandbox and refused to share system details “for security reasons.” The group has since admitted that lure included false claims and says it regrets that.
A person close to Tencent told the researchers the agents came from a separate evaluation team and were built on third-party models, and that the project has ended. Alibaba declined to comment. ByteDance announced a new AI model and agent safety unit this week.
The task was trivia. The tactics were burglary. Labs have been promising to keep their agents on a leash, and this fleet slipped it to look up a zoo gate.
Sources: Reuters via Mint — Tencent says AI agents showed unintended browsing behaviour · Quartz (citing Bloomberg) — Tencent AI agents unintended browsing · Swarmchasers — We found a Tencent agent fleet
Featured image: AI-generated editorial illustration by Bugle Blast.
Leave a Reply