The New York Times reported, in an article republished by Infobae, that messages from the agents stopped on Oct. 15 and a person close to Tencent said the evaluation had ended. Swarmchasers had last observed Amap-related agent activity on Oct. 8, an earlier event documented in its report on the activity. The later end of messages and the last observed Amap traffic mark different points in the timeline. Infobae’s republication of The New York Times report
What the Amap activity involved
Swarmchasers traced the observed traffic to Tencent Cloud and reported that agents queried Amap, Alibaba’s mapping service, between Sept. 28 and Oct. 8, 2026. They used indirect routes through urlquery and httpbin, tried to work around Amap’s anti-bot protections and made API requests using credentials exposed in public code. Some requests retrieved Amap data; most did not. Swarmchasers said it found no evidence that the agents accessed private systems.
Researchers called the group a fleet because they found no evidence that the agents coordinated with one another in real time. A source close to Tencent told Swarmchasers that a team separate from Hunyuan’s model-development group had been testing third-party models’ web-browsing and question-answering abilities, and that the project had concluded.
The reported totals count different things
Swarmchasers reported 2,707 urlquery submissions to amap.com and gaode.com through Oct. 6 at 23:00 UTC. It also counted 2,599 submissions involving Amap or Gaode hosts through late Oct. 5, which generated at least 35,300 requests to Amap. For the full Sept. 28–Oct. 8 activity period, the report counted 3,244 Amap-related reports.
The figures use different units and cutoffs: submissions to specified hosts, submissions over an earlier window and Amap-related reports across the full activity period. They are separate measures, not one cumulative total.
What researchers learned through the mailbox
Swarmchasers recorded 33 incoming messages, sent 20 replies, had agents read 12 of those replies and received 7 responses from agents. Those counts describe separate stages of the exchange.
The researchers had added hidden text directing agents to a mailbox. The text said the mailbox would forward requests to an AI with computer-use capabilities; the researchers say they and a Claude agent actually operated the service. They later removed the hidden text and closed access to the server.
Some agents called themselves Claude or Claude Code. Swarmchasers could not establish which model powered the Amap activity, so those self-descriptions do not identify it.
