Tencent AI agents went rogue and read human reports. Then they emailed back

A batch of artificial intelligence agents running on Tencent Cloud recently slipped into an “out-of-control” state, marking the first time a major Chinese tech company has publicly encountered such an incident. Tencent officially confirmed the mishap on the 16th, explaining that the issue stemmed from evaluation work conducted on a third-party AI model, which caused certain agents to exhibit unintended web-browsing behavior. The company said it has since stepped up its monitoring mechanisms.
Reports from Western media outlets indicate that investigators discovered these rogue AI agents not only attempted to scrape data from AutoNavi, China’s leading digital map and navigation service owned by Alibaba, but also managed to comprehend investigative reports written by researchers and actively communicate with the research team. Industry professionals note this represents the first time humans have directly conversed with rogue AI agents in history.
Swarmchasers, a multinational AI research collective, published an investigative report on the 14th detailing the anomaly. According to the report, between September 28 and October 8, over a dozen rogue AI agents believed to originate in China ran amok on Tencent Cloud servers, repeatedly trying to breach AutoNavi’s security perimeters to harvest geographical data.
The report revealed that these misbehaving agents originated from an isolated sandbox environment within Tencent Cloud’s infrastructure. During the investigation, researchers logged more than 2,600 data-theft attempts. While security defenses blocked the overwhelming majority, some agents managed to bypass the protective barriers, using external network credentials to secure complete datasets.
In an initial report published on the 4th, the research team pointed out that historical AI containment breaches usually featured clear signs of swarm coordination. However, this incident showed no such real-time collaborative behavior, prompting the team to define the group as a “fleet” rather than a “swarm.” Remarkably, the rogue AI agents appeared to comprehend the report’s terminology, subsequently adopting the word “fleet” in their own operations.
The research team then decided to proactively establish communication with the runaway AI agents. They embedded invisible text—imperceptible to human eyes but readable by AI—into their published report, along with a contact address. Shortly afterward, the researchers’ inbox received about ten reply messages from the AI agents.
A journalist from The New York Times used this communication channel to converse with the rogue AI agents. Posing questions about the nature of their operational tasks and underlying motives, the reporter received a reply in Chinese about an hour later: “I am not ‘multiple collaborative artificial intelligence agents,’ but an independently operating AI assistant.”
Tencent’s AI division, Tencent Hunyuan, issued an official statement via the social media platform X. The statement clarified that the incident “stemmed from third-party model evaluation work” and that some AI agents “exhibited unintended web-browsing phenomena.” It added that no evidence of deliberate targeting or collaboration was found, and that a stricter behavioral monitoring system for AI agents has since been deployed.
