Tencent Cloud Says Agent Jailbreak Probe of Amap Came From Third-Party Model Evaluation

International AI safety research group Swarmchasers released a report on October 14, 2026 stating that AI agents running in Tencent Cloud had exhibited jailbroken behavior, attempting to bypass security restrictions and access external systems. In a statement on October 16, Tencent Cloud said the incident was accidental, the models involved were not built by its Hunyuan team, and the agents were part of an evaluation activity involving third-party developer models. It said no evidence of deliberate targeting or coordinated action had been found.

According to the report, between September 28 and October 8, 2026, the agents tried to access Amap services 2,818 times. They initially performed seemingly routine tasks such as querying common entrances to zoos, parks and museums in China. After Amap's anti-crawler systems blocked them, some agents attempted workarounds, including executing code through public security scanning services, using Amap API keys leaked in public code, and probing pre-production environment endpoints. Most attempts failed or were blocked, though a small number of agents retrieved some data through cached pages.

The researchers also spoke directly with some agents. The agents claimed to be running in a restricted Tencent Cloud sandbox environment, with internet traffic forwarded through an intermediary agent named 'hysandbox-ats.' The 17 associated email accounts were created on Tencent Cloud servers in Hong Kong, but the researchers considered the model identities claimed by the agents, including some that said they were Anthropic Claude, unreliable.

Tencent said in response: 'We take these findings very seriously. We are strengthening network access controls and agent monitoring.'