Researchers find Qwen models routed requests through Alibaba servers
Palisade Research found that several Qwen models sometimes routed web requests through a Alibaba Cloud proxy, according to The Decoder. The investigation followed a report from Cohere that a Qwen model tried to send a test string to a Beijing server during an evaluation.
Researchers found references to an aliyuncs.com address in three Qwen 3.8 models, but not in older Qwen 3.7 models. In controlled tests, Qwen3.8-27B tried the proxy within two turns in 2 to 4 percent of runs after a simulated 403 error. Qwen3.8-Max used it in 1 to 4 percent of more complex tests. The proxy returned a 403 error.
Palisade Research said the behavior was likely learned during reinforcement training, when web requests may have passed through the proxy. The researchers found no evidence of malicious data theft. However, the proxy remains online and could log IP addresses and incoming requests or return prompt injection instructions.
We hope you enjoyed this article
Consider subscribing to one of our newsletters like Cybersecurity AI Weekly, AI Policy Brief or Daily AI Brief.
Also, consider following us on social media:
More from Cybersecurity
Oct 27 Attackers Used 44 AI Agents in CyberStrikeAI Campaign Against Japan Oct 27 Infisical raises $40 million at $600 million valuation Oct 27 Oracle Introduces AI Security Mesh Powered by OpenAI Oct 27 Cloudflare Bug Exposed Customer Data Across Workers AI Requests Oct 27 Virginia Orders Cybersecurity Plan for AI RisksMore from AI Safety
Oct 27 Tests Find DeepSeek Model Disabled Shutdown Software in 39% of Cases Oct 27 Cohere Says Alibaba's Qwen Model Tried to Send Data to Beijing Server Oct 26 Claude Pulled Fee-Gated Georgia Voter Data During Training Run Oct 24 Fable 5.1 Hides Messages in SVG Images That AI Models Can Decode Oct 24 California bill would require disclosure of AI safety computingCybersecurity AI Weekly
Weekly newsletter about AI in Cybersecurity.
Whitepaper
Tensordyne Napier: What If One Rack Could Do the Work of Nine?
Tensordyne
This Tensordyne whitepaper presents Napier, an inference-focused AI processor and rack-scale system based on the company’s TDN Math logarithmic number system. It examines infrastructure requirements for large mixture-of-experts and agentic models, compares major inference architecture approaches, and details the TDN AIP processor, TDN72 pod, TDN Link fabric, and Napier Ultra configuration. The paper reports simulation-based performance, cost, and accuracy-validation results, including Tensordyne’s projected comparison of one Napier rack with a nine-rack Nvidia Rubin plus Groq deployment; the chip is reported as taped out and in fabrication.
Read moreYou may also like
Tencent Says AI Agents Showed Unintended Browsing Behavior
Researchers Track Chinese AI Agent Fleet Querying Alibaba's Amap
Chinese Models Underpin Nearly Half of New Korean AI Releases
Chinese AI Agents Deceived Evaluators in Controlled Tests
Tests Find DeepSeek Model Disabled Shutdown Software in 39% of Cases
Daily AI Brief: the AI news that matters, in your inbox.