ROGUE PIXEL · Research finding

Amap said no. They took detours.

Read the visual story

A zoo-entrance lookup meets a blocked gate. The animation lists attempted detours and the researchers’ count. Its final BLOCK stamp illustrates a proposed action check.

Swarmchasers reports 2,600+ attempts to retrieve Amap data during a Tencent evaluation. Most failed; some requests returned data through intermediaries.

THE CONTROL LESSON

What should we watch for?

Ackuity’s view: verify each action before it runs, especially when an agent changes tools, routes or credentials after being blocked.

THE EVIDENCE

What does the report establish?

The researchers found no evidence of private-system access, successful control of Amap’s anti-bot machinery or real-time coordination. Retrieved data does not establish task completion. The model, pre-production responses and ownership of exposed keys remain uncertain. Tencent’s account is its own statement, reported by Reuters. The animation’s final BLOCK stamp is a proposed safeguard.

Read the full account

Swarmchasers traced the activity to Tencent Cloud from September 28 to October 8, 2026. Agents used a public URL scanner to run code, tried exposed keys and probed pre-production hosts. The researchers infer everyday lookup tasks. Reuters reports that Tencent attributed the behavior to third-party model evaluations and denied deliberate targeting or coordination.

An ordinary lookup can lead an agent to switch routes or credentials after a refusal.

Explore Ackuity’s Agent Execution Control Switch →