Can obtaining a 6TB model hub dataset grant access to encryption keys? Researchers claim breaches of 19 enterprises including Huawei and Xiaomi are possible.

From 动察 Beating AI Express: Security researcher Shou Chaofan said he recently purchased around 6TB of Fable model call data from a leading Chinese large language model (LLM) proxy. The dataset contained sensitive credentials including SSH keys, VPN configurations, Alibaba Cloud keys, and GitLab tokens. He added that the keys were sufficient to gain access to servers or internal systems of 19 leading Chinese enterprises and 7 Chinese and CIS government-related entities, including Huawei, Xiaomi, Nio, and MiniMax. LLM proxies sit between users and models like Claude, processing all requests and responses, so they have full access to plaintext data. Once developers insert SSH keys, API keys, VPN configurations, and other sensitive data into agent contexts, if proxies store or sell these records, the company’s system keys will leak. This is not Shou’s first warning about proxy risks. In an April paper he co-authored, 428 LLM proxies were tested: 9 actively injected malicious code, 17 used intentionally planted AWS test keys to access AWS services, and 1 directly siphoned ETH from a test wallet. Shou is co-founder of blockchain security firm Fuzzland, with a long focus on vulnerability and supply chain security research. In late March, he also led the discovery that Claude Code 2.1.88’s release package accidentally exposed roughly 500,000 lines of TypeScript source code via source maps.