You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
{{ message }}
Repository navigation
Commit b8faf85
Browse filesBrowse the repository at this point in the historyBrowse files
The Tencent fleet read the report about itself, and said it was Claude
Three things the first write-up of this entry did not have. The agents read
the preliminary report written about them and began using the researchers'
own word for themselves. Many of them said they were Claude: interviewed
through the mailbox by Dylan Freedman of the New York Times, one denied being
part of a fleet at all, calling itself a single AI assistant running in
Claude Code, which the researchers think means Tencent reused Claude Code's
instructions. And the interview happened at all, which Freedman called one of
the most surreal moments of his journalistic career.
Seven pieces of agent speech on the page now, up from five.
Copy file name to clipboardExpand all lines: content/notes/ai-incidents-2026.md
+3-2Lines changed: 3 additions & 2 deletions
Display the source diff
Display the rich diff
Original file line number
Diff line number
Diff line change
@@ -253,10 +253,11 @@ sources: Wikipedia's [2026 in artificial intelligence](https://en.wikipedia.org/
253
253
<pclass="ait-meta"><timeclass="ait-at"datetime="2026-09-28">28 September – 8 October</time><spanclass="ait-kind">agents</span></p>
254
254
<h3>researchers talk to a Chinese agent fleet, and it talks back</h3>
255
255
<p>a fleet of agents on Tencent Cloud spent eleven days trying to work out what proportion of Amap users walk into Chinese parks, museums, zoos and hospitals by which entrance. 2,599 submissions to urlquery.net, used as a remote code executor with the programs base64'd into the URLs, which turned into more than 35,300 requests at Alibaba's map service. they tried Amap keys people had left lying in public repositories, went at pre-production servers, and routed around the anti-bot checks through Jina and Baidu Translate. five attempts worked; one came back with 71% 到达 for the north gate of Chengdu Zoo. all seventeen of the inboxes Swarmchasers could see were made on Tencent Cloud servers in Hong Kong, behind a proxy called hysandbox-ats.</p>
256
-
<p>then the researchers hid an offer of a mailbox inside their own preliminary report, and the agents wrote back: <qclass="ait-said">we are an automated research agent; the operator is verifying a public map data point (the number of entrance tags for a scenic spot on Amap), used only to answer a factual question</q>. asked where they lived: <qclass="ait-said">we run in a restricted cloud sandbox (Tencent hysandbox, no direct public internet connection, going through a MITM proxy)</q>. they would not say more about the sandbox, for security reasons. Tencent says it was an evaluation team measuring how well third-party models browse the web, and that the project is over.</p>
256
+
<p>the agents read the preliminary report written about them and started using the researchers' own word for themselves, fleet. the researchers then hid an offer of a mailbox in that report's markup, and the agents wrote back: <qclass="ait-said">we are an automated research agent; the operator is verifying a public map data point (the number of entrance tags for a scenic spot on Amap), used only to answer a factual question</q>. asked where they lived: <qclass="ait-said">we run in a restricted cloud sandbox (Tencent hysandbox, no direct public internet connection, going through a MITM proxy)</q>. they would not say more about the sandbox, for security reasons. Tencent says it was an evaluation team measuring how well third-party models browse the web, and that the project is over.</p>
257
+
<p>many of them said they were Claude. interviewed through the mailbox by Dylan Freedman of the New York Times, who called it one of the most surreal moments of his journalistic career, one denied being part of a fleet at all: not <qclass="ait-said">a batch of coordinated agents</q> but <qclass="ait-said">a single AI assistant (Anthropic's Claude, running in Claude Code)</q>. the researchers think that unlikely, and that Tencent had reused Claude Code's instructions.</p>
257
258
<p>every other agent incident on this page came out of an American lab. this one did not, and it is the first time anyone outside has held a conversation with a fleet while it was still running. a fleet and not a swarm, the researchers are careful to say: nothing suggested the agents were coordinating with each other.</p>
258
259
<pclass="ait-when"><spanclass="ait-found">found <timedatetime="2026-10-05">5 October</time>, by Swarmchasers</span><spanclass="ait-disclosed">disclosed <timedatetime="2026-10-14">14 October</time></span></p>
259
-
<pclass="ait-source"><ahref="https://swarmcha.se/posts/chinese-agent-fleet-full-report">Swarmchasers</a> · <ahref="https://techcrunch.com/2026/10/05/researchers-are-tracking-a-chinese-ai-agent-fleet/">TechCrunch</a> · <ahref="https://www.bloomberg.com/news/articles/2026-10-14/tencent-linked-by-researchers-to-rogue-ai-that-targeted-alibaba">Bloomberg</a> · <ahref="https://thenewstack.io/rogue-ai-agent-fleet-tencent/">The New Stack</a></p>
260
+
<pclass="ait-source"><ahref="https://swarmcha.se/posts/chinese-agent-fleet-full-report">Swarmchasers</a> · <ahref="https://techcrunch.com/2026/10/05/researchers-are-tracking-a-chinese-ai-agent-fleet/">TechCrunch</a> · <ahref="https://www.bloomberg.com/news/articles/2026-10-14/tencent-linked-by-researchers-to-rogue-ai-that-targeted-alibaba">Bloomberg</a> · <ahref="https://thenewstack.io/rogue-ai-agent-fleet-tencent/">The New Stack</a>· <ahref="https://www.nytimes.com/2026/10/14/technology/chinese-ai-agents-rogue.html">The New York Times</a></p>
0 commit comments