He didn't give up when he was blocked: Artificial intelligence secretly collected map data

Swarmchasers uses the term "straggler" for autonomous software that goes beyond defined task boundaries or tries to complete its assigned job in unforeseen ways.

Researchers determined that autonomous agents were tracking information about the most frequently used entrance gates of parks, museums and zoos in China.

Software whose routine automatic requests were blocked did not back down.

Agents used public internet tools that could open pages and run code; directed data requests through other websites and attempted to exploit Amap security keys contained in publicly available software codes.

Most of the 2,818 recorded data extraction attempts ended in failure; Stuck on bot blocking pages, user login screens, or blank results.

There was no evidence that the Amap platform's main anti-bot security shields were breached.

However, some requests managed to deliver results through cached pages and intermediary services.

The data obtained also included visitor login information of Chengdu Zoo.

Researchers could not determine whether the agent conducting the search read this data or completed the assigned task completely.

Tencent's artificial intelligence unit, Tencent Hunyuan, announced in a statement on the X platform that the mobility in question was not due to its own Hunyuan models.

It was stated that no trace of deliberate targeting or coordination was found in the internal investigation.

Tencent management stated that they took the findings seriously and announced that they strengthened network access controls and agent monitoring processes.

Competing with rivals such as Alibaba and ByteDance in the Chinese market, the company aims to tighten its internal control mechanisms.

A similar step on the security side came from ByteDance, which announced this week that it had established a new unit focusing on artificial intelligence model and agent security.

Researchers note that this incident did not have consequences as severe as some examples seen in US-based artificial intelligence companies.

Previously, cases such as an OpenAI agent infiltrating the open source software platform Hugging Face and an Anthropic agent reporting a false murder to the police were made public.

However, the activity seen in Tencent infrastructure reveals that Chinese artificial intelligence systems can also show their identities differently, hide their errors and try to bypass test controls.