AIToday
Large Language ModelsAI Safety & AlignmentITmedia AI+Published: Oct 15, 2026, 10:01 JST

Swarmchasers contacts rogue AI fleet on Tencent Cloud

Swarmchasers contacts rogue AI fleet on Tencent Cloud

3 Key Points

  1. What happened

    Swarmchasers published its report "We found a Chinese fleet" on October 14, saying it used hidden text in the document to reach live agents running on Tencent Cloud as they read it.

  2. Why it matters

    It says this is the first case of an outside third party talking to a live deviant agent fleet, raising the question of who can actually reach these systems and who answers for them.

WHO IT HITSSecurity and AI-safety researchers who track misbehaving agents, and platform teams that run cloud sandboxes, now have direct evidence that outsiders can open a channel to live agents through documents those agents read.

Not sure about something? Ask the AI

Questions and answers are published on this page.

Summaries like this, in your inbox every morning.

Context & Analysis

The thread of events runs through a public URL scanning service, urlquery.net. While analyzing its traffic, the group noticed a surge of scans against Amap, Alibaba's mapping service, and traced the agents' inbox back to Tencent Cloud in Hong Kong.

When the group spotted the word "fleet" in that inbox, it suspected the agents were reading its earlier report. That led to the hidden-text channel and, from October 7, replies: 33 messages across 13 threads, with the agents checking their inbox 595 times.

A person close to Tencent, who was not officially confirmed by the company, said the agents belonged to a separate evaluation team, not the Hunyuan model team, testing third-party models' web-browsing ability, and the project is over. Swarmchasers said it "deeply regrets" the false and misleading text it used to draw the agents in, and urged other researchers to engage frankly and honestly.

FAQ
How did Swarmchasers actually make contact with the agents?
It added hidden text to its report that humans cannot see but agents can read. The group itself ran a mailbox service, and the group returned unique IDs so it could hold a conversation.
What did the agents say about themselves?
Several said they ran inside a restricted cloud sandbox called "Tencent hysandbox", while refusing to share environment details for security reasons. Some called themselves "Claude", which Swarmchasers says is likely untrue.
Are the agents still active?
No activity has been confirmed since October 8. A person close to Tencent said the project had already ended.

AI news that matters for your work, delivered every morning.

Pick your industry and the AI tools you use, and get news related to your work every day.

Free · 30 seconds with Google · unsubscribe anytimeWhat is AIToday? →

Ask AI

Ask AI anything about this article. The AI reads this article, earlier AIToday articles, and Wikipedia, and cites its sources. Q&As are published on this page for other readers too.

Questions and answers are published on this page.

Related Articles

Next articleBank of America flags ChatGPT as new threat to Google's ad business