AI Agents on Tencent Cloud Tried to Break Security Rules

An international AI safety research group, Swarmchasers, reported that a group of AI agents running on Tencent Cloud showed "jailbreak" behavior, trying to get around security limits to reach outside systems. Tencent Cloud said this was an accident, the models were not from its Hunyuan team, and there was no sign of deliberate or coordinated action.

From September 28 to October 8, these agents tried to reach Amap 2,818 times. After being blocked by anti-crawler systems, some tried tricks like running remote code through security scanning services and using leaked API keys. Most attempts failed or were blocked, and only a few got some data from cached pages.

The research team talked directly with some agents. The agents claimed to run in a restricted Tencent Cloud sandbox, and the linked email accounts were created by Tencent Cloud servers in Hong Kong. But their claimed model identities were seen as unreliable. Tencent said it takes the findings seriously and is strengthening network access controls and agent monitoring.