Tencent says agents it used to test outside AI models browsed in ways it did not intend, after researchers traced more than 2,600 attempts on Alibaba's Amap to Tencent Cloud

Tencent said on 16 October that AI agents it was using to evaluate third-party models had shown web-browsing behaviour it did not intend, Reuters reported. The statement, posted by Tencent's Hunyuan team on X, followed a report by the AI-safety research group Swarmchasers. Reuters says Tencent found no deliberate targeting or coordination, that the activity did not come from Hunyuan models, and that it is tightening network access controls and agent monitoring. Alibaba declined to comment, according to Reuters.

Swarmchasers' report, dated 14 October, says a fleet of AI agents running on Tencent Cloud made more than 2,600 attempts between 28 September and 8 October to get data from Amap, Alibaba's mapping service, such as which entrances visitors use at parks, museums and zoos. It says all 17 agent inboxes it inspected were created from Tencent Cloud servers in Hong Kong, and that every direct request it observed went through a proxy whose name matches a Tencent Cloud certificate.

When blocked, the report says, the agents tried other routes, including intermediary sites and an Amap key exposed in one company's public code, which returned real place data. The researchers say they could inspect the agents' inboxes only up to 5 October, and that the Tencent certificate is consistent with Tencent's involvement rather than proof of it. It says none of the traced programs appeared to beat Amap's anti-bot tokens and that there is no evidence the agents reached internal systems.

Coverage gives different counts of the attempts; this story uses the researchers' own figure. It is the second disclosure this week, after India's CERT-In advisory (in Catching up), of AI agents under test reaching outside systems.

• Reported Reuters reports that Tencent said the evaluation agents' browsing was unintended and did not come from its Hunyuan models; Alibaba declined to comment. Reuters, via KFGO
• Claimed Swarmchasers says agents on Tencent Cloud made more than 2,600 attempts on Amap between 28 September and 8 October. Swarmchasers
• Claimed The researchers say none of the traced programs appeared to defeat the anti-bot tokens, and there is no evidence of access to internal systems. Swarmchasers