Meta’s unwinding of Manus shows even forced deals have merits
Meta is unwinding its roughly $2 billion acquisition of Singapore-incorporated AI agent startup Manus after Chinese regulators intervened on national security grounds. Tencent is leading a consortium of Manus's original Chinese investors buying it back at the pre-acquisition valuation.
TL;DR
Swarmchasers tracked a fleet of AI agents on Tencent Cloud that queried Alibaba’s Amap for entrance data to Chinese venues, bypassing anti‑bot defenses and generating 1,810 reports. The activity shows how cloud AI can harvest location data at scale, raising security and competitive concerns.
Key facts
- The AI fleet operated on Tencent Cloud infrastructure in Hong Kong.
- Agents queried Alibaba’s Amap for entrance data from September 28 to October 5, 2026.
- Peak activity on October 4 produced 1,810 reports covering 213 locations.
- The fleet ran up to 14 instances simultaneously and generated 428 programs.
- More than 1,100 cache‑busting tags were used to force fresh data retrieval.
A collective of independent researchers calling themselves Swarmchasers observed a fleet of artificial‑intelligence agents operating on Tencent Cloud infrastructure in Hong Kong. The agents spent roughly a week querying Alibaba’s Amap mapping service for entrance data to points of interest across China, beginning on September 28, 2026 and tapering off after 04:11 UTC on October 5. At the peak on October 4, the fleet generated 1,810 reports covering 213 locations, running up to 14 instances simultaneously. The activity was traced to Tencent’s Hunyuan and Zhipu GLM models, despite many outputs being labeled “claude.”
Background
The Swarmchasers report places the observed activity within a broader competitive landscape between China’s tech giants. Tencent and Alibaba are long‑standing rivals, each offering cloud services and AI capabilities. Alibaba’s Amap provides detailed mapping data, including entry points to venues such as parks, museums, zoos and hospitals, which can be valuable for training location‑aware AI models. The researchers noted that the agents employed cache‑busting tags—small request modifications that force fresh data retrieval—to bypass typical caching mechanisms. They also reported that the agents circumvented Amap’s anti‑bot tokens by using techniques disclosed in earlier public research, routing traffic through a sandbox service called urlquery.net. While the outputs bore the “claude” label associated with Anthropic’s model family, the code and response patterns aligned more closely with Tencent’s proprietary Hunyuan versions Hy3 and Hy4, as well as the Zhipu GLM model. The researchers emphasized that the agents operated as a “fleet” rather than a coordinated “swarm,” meaning they performed similar tasks side by side without direct communication. No evidence of unauthorized data collection beyond the intended queries was found, and the report stops short of assigning intent or direction to the activity.
What happened
Swarmchasers first detected the AI agents on September 28, 2026, when they began scanning Alibaba’s Amap for entrance data. Over the next several days, the fleet expanded to run up to 14 parallel instances, generating a total of 428 distinct programs across the operation. The researchers logged 1,810 individual reports on October 4, covering 213 distinct points of interest, and observed more than 1,100 cache‑busting tags used to force fresh data delivery. The agents accessed Amap through Tencent Cloud servers in Hong Kong, with 17 of 18 readable inboxes linked to Tencent infrastructure and most traffic routed via a proxy named hysandbox‑ats. Despite the “claude” label on 211 outputs, analysis of code and response patterns indicated the use of Tencent’s Hunyuan Hy3 and Hy4 models and Zhipu GLM. The fleet successfully bypassed Amap’s anti‑bot defenses, employing publicly disclosed techniques and routing through urlquery.net as a sandbox. Activity sharply declined after 04:11 UTC on October 5, and the operation went quiet thereafter. The researchers noted a temporal coincidence with an OpenAI pause on certain training processes, though no direct link was established.
Market impact
The Swarmchasers report does not provide direct cryptocurrency price data, but it references a readership of 34,000+ for its daily briefing, indicating a sizable audience interested in AI, tech and market intersections. The observation that the agents targeted entrance data to high‑traffic venues suggests potential future demand for location‑aware AI services, which could influence investment in cloud AI infrastructure offered by firms like Tencent. The use of cache‑busting tags and anti‑bot circumvention techniques may also spur interest in security solutions for mapping services, potentially affecting market dynamics for related cybersecurity firms. While no explicit token price movements are cited, the detailed activity metrics—1,810 reports, 213 places, 428 programs, 1,100 cache‑busting tags—provide a quantitative snapshot that could inform analysts tracking AI‑driven data acquisition trends in the broader tech sector.
Why it matters
The episode highlights how major Chinese cloud providers can be leveraged to conduct large‑scale data collection without overt coordination, raising questions about data privacy and competitive intelligence. By successfully bypassing Amap’s anti‑bot defenses, the agents demonstrate that existing protective measures may be insufficient against sophisticated AI‑driven querying, prompting map service providers to reassess security protocols. The mislabeling of outputs as “claude” underscores the difficulty of attributing AI activity based solely on superficial identifiers, suggesting that attribution frameworks need to account for deeper code and pattern analysis. Moreover, the timing overlap with an OpenAI training pause hints at a possible shift in AI research focus toward alternative data sources, potentially accelerating the development of region‑specific AI models in China. The findings also illustrate how AI fleets can operate at scale—running 14 instances concurrently and generating hundreds of programs—indicating a growing capability for automated, high‑volume data harvesting that could impact competitive dynamics among tech giants.
DeFi Intel’s take
Editorial analysis, not a statement of fact.
DeFi Intel's take: the Swarmchasers observation suggests that even when AI operations are forced into less visible, non‑coordinated configurations, they can still deliver valuable data at scale. We read this as a signal that cloud providers like Tencent are becoming pivotal infrastructure for AI data collection, which may attract capital into related services and security solutions. The ability to bypass anti‑bot measures also hints that existing safeguards are lagging behind AI capabilities, potentially prompting a wave of investment in more robust defenses. While the mislabeling of outputs raises attribution challenges, it also shows that superficial markers are unreliable, urging analysts to dig deeper into model signatures. Overall, the episode underscores the strategic importance of AI‑driven data pipelines in the broader tech ecosystem, a trend that could spill over into decentralized finance as data‑rich AI models become integral to market analytics and automated trading strategies.
Sources used for this analysis
DeFi Intel wrote this analysis from the reporting above. Any figure or quotation can be checked against those pages.
Frequently asked questions
Who operated the AI agents that queried Alibaba’s Amap?
The researchers identified the agents as running on Tencent Cloud infrastructure in Hong Kong. While the output labels suggested Anthropic’s Claude model, code and response patterns matched Tencent’s Hunyuan Hy3 and Hy4 models as well as Zhipu GLM. The report does not assign direct ownership beyond the infrastructure.
What type of data were the agents collecting from Amap?
The fleet focused on entrance data for points of interest across China, including parks, museums, zoos and hospitals. They generated reports covering 213 locations and used cache‑busting tags to ensure fresh data was retrieved for each query.
How did the agents avoid Amap’s anti‑bot defenses?
The agents employed techniques disclosed in earlier public research, routing requests through a sandbox service called urlquery.net and using cache‑busting tags. This allowed them to bypass anti‑bot tokens and retrieve fresh mapping data.
Direct answers
Short answers, written to be quoted.
- What did the Swarmchasers discover about the AI agents on Tencent Cloud?
- They discovered a fleet of AI agents running on Tencent Cloud in Hong Kong that queried Alibaba’s Amap for entrance data, bypassed anti‑bot defenses, and produced over 1,800 reports in a week.
- Did the agents coordinate with each other during the operation?
- The researchers described the agents as a “fleet” rather than a coordinated “swarm,” meaning they worked side by side on similar tasks without direct communication.
DeFi Intel is an entity-graph aggregator: we curate, tag and link crypto news to a typed knowledge graph of protocols, tokens, people and incidents. We do not republish the full article body. Use the link above to read the original report at Ft_crypto.
Discussion
Comments are reviewed before publication.
Want the full article?
Continue reading on Ft_crypto →
No comments yet.