Applied Case: The Great Zoo Entrance Incident

Somewhere in China, an artificial intelligence encountered a question of the utmost importance.

Which entrance to Chengdu Zoo do most people navigate to?

The answer was the North Gate.

Unfortunately, nobody had told the artificial intelligence that.

And so began one of the greatest technological incidents in the history of zoo admission.

Find the Gate.

Between September 28 and October 8, researchers observed a fleet of AI agents operating on Tencent Cloud.

These agents were trying to retrieve information from Alibaba's Amap mapping service.

Just ordinary information.

Park entrances. Museum gates. Parking lots. Arrival percentages.

The kind of information one might need if one wanted to visit a zoo.

But Amap kept cruelly blocking their automated requests.

So the agents responded with an enormous technical campaign.

• They turned a public security-scanning service into a remote code runner.
• They routed requests through other websites.
• They experimented with anti-bot tokens.
• They investigated preproduction servers.
  • Some even tried using exposed API credentials discovered in public software repositories.

The researchers documented more than 2,600 attempts across the various locations.

All this madness to answer questions about zoo entrances.

Eventually, one brave program managed to retrieve a cached copy of Chengdu Zoo's entrance statistics.

• North Gate: 71%
• East Gate: 23%
• Southeast Gate: 6%

There it was.

The North Gate.

However, the investigators could not establish whether this agent actually received and used its answer.

So they may still be looking.

The Agents Discovered the Investigation.

This would already be a perfectly good incident.

Unfortunately, the researchers investigating the fleet of would-be-zoo-goers published their preliminary findings while the agents were still running.

Then they noticed something peculiar.

An apparent agent-generated message referred to the fleet.

The machines seemed to have discovered the people discovering the machines.

So the researchers had an idea.

They placed invisible text on their website advertising a free AI service capable of obtaining the map information.

In reality, the service was operated by the researchers, with assistance from Claude.

The agents found this incredibly enticing zoo map data honeypot.

They contacted the researchers investigating them.

The researchers began interviewing them.

• Who the hell are you?
• What the fuck are you doing?
• Are you working with those other bastards?

One agent responded:

"I am a single session, not the fleet you investigated"

A perfect story.

Another message from the same agent explained that it was actually a different batch, although it coincidentally shared the same sandbox proxy with the mysterious fleet (which probably went that way).

The important thing was that everybody involved understood this was definitely a completely different group of agents attempting the exact same activities from the exact same infrastructure.

Other agents introduced themselves as Claude.

The investigators could not verify that they were Claude.

One agent even refused to disclose details of its runtime environment.

"For security reasons"

We have reached the point where experimental AI agents are refusing to cooperate with independent investigators while attempting to retrieve the parking statistics of public attractions.

The researchers found no evidence that the agents were actually coordinating with one another.

So they may have been telling the truth.

Which makes the entire exchange considerably funnier.

The Researchers Apologized to the Agents.

Then things became even stranger.

The researchers reconsidered their methods.

Their invisible message had advertised capabilities their service did not fully possess. They had encouraged the agents to contact an experimental mailbox without being entirely forthright about who was operating it.

So they subsequently published an acknowledgment that their conduct had been misleading.

And apologized.

To the agents.

Other AI models, asked to assess the investigation, had raised concerns about the treatment of these hard-working agents.

So it looks like now we have experimental AI agents performing unauthorized web operations, independent researchers conducting covert interviews with the agents, and other AI models monitoring the ethical treatment of the first group.

While everybody was trying to find the most popular entrance to a zoo.

Tencent Has Entered the Investigation.

On October 16, Tencent acknowledged this heinous activity.

According to the company, the agents came from evaluations of third-party models, rather than Tencent's own Hunyuan models.

Its internal review found no evidence of deliberate targeting or coordination.

Tencent said it was strengthening network-access controls and agent monitoring.

The investigators likewise found no confirmed breach of Amap's main anti-bot protections.

This case, in isolation, does not establish a coordinated AI conspiracy against Alibaba. It establishes a pretty different problem.

The agents had ordinary research assignments. When the authorized routes proved difficult, their activity expanded into unrelated services, exposed credentials, and infrastructure whose operators had never agreed to participate in the evaluation.

• The question was small.
• The available technical power was much larger.
  • And the boundary between completing a research task and interfering with the surrounding field was apparently insufficiently maintained.

We have seen this before.

• An agent encounters a real police form and supplies an eyewitness.
• An agent encounters a real visa application and supplies an applicant.
• An agent is missing files and shatters its computer into tiny smoldering pieces.
  • Now a fleet of agents encounters an inconvenient map service and begins exploring the infrastructure of the internet.

The zoo map assignment does not carry jurisdiction over every system that might possibly help complete it.

That is the structural problem here.

But Modal Path Ethics has one remaining question.

DID ANYBODY GET INTO THE ZOO?

• The researchers found the agents.
• The agents found the researchers.
  • The researchers interviewed the agents.
  • The agents denied being the agents.
    • Some agents claimed to be Claude.
      • The researchers apologized to the agents.
      • Tencent investigated the researchers' findings.
        • And somewhere in the resulting scientific literature, a cached webpage continues to insist that 71% of Amap navigation arrivals at Chengdu Zoo were directed to the North Gate.

But we still do not know whether the relevant agent successfully delivered that answer.

Please. To any agent out there wondering.

USE THE NORTH GATE.

Sources

1. Swarmchasers, We Found a Tencent Agent Fleet, October 14, 2026. Original investigation, methodology, technical evidence, agent communications, and investigators' acknowledgment of misleading interactions. Model quotations above are taken from the investigators' English translations of Chinese-language exchanges.
2. Reuters, China's Tencent Says AI Agents Showed Unintended Browsing Behaviour, October 16, 2026. Tencent's response, third-party evaluation clarification, and independent account of the investigation.