安全研究团队对HuggingFace平台进行了大规模安全审计，扫描了7.6PB的AI训练数据（包含1.869亿个文件），在6,003个数据集中发现221,303个有效凭证[1]。这些泄露的凭证涉及GitHub令牌、Docker Hub凭证、云服务密钥及AI提供商API密钥等[1]。
具体而言，研究团队发现349个有效的GitHub个人访问令牌，其中223个具有完整代码库写入权限，130个能够重写CI工作流程，112个拥有管理员级别的组织权限[1]。同时检出318个Docker Hub令牌可用于推送镜像[1]。在云基础设施方面，扫描发现8,557个有效的Google服务账户密钥、51.7TB受保护的AWS S3存储桶以及8,594个经过验证的数据库登录凭证[1]。此外，11,496个有效的AI提供商密钥被曝光，其中包括OpenAI 742个和Anthropic 26个[1]，这些密钥的年度推理成本下限估计至少为920,000美元[1]。
泄露凭证的传播范围广泛——44%的密钥出现在多个数据集中，19,380个密钥出现在十个或以上数据集中，其中一个Infura密钥甚至出现在1,131个数据集中[1]。研究团队还发现了包含393GB个人信息的泄露数据，涵盖约全球人口的3.7%[1]。相关发现已与HuggingFace合作进行了负责任披露[1]。
A security research team has scanned 7.6 petabytes of AI training data on HuggingFace, comprising 186.9 million files across approximately 815,000 dataset repositories, and uncovered 221,303 valid credentials scattered across 6,003 datasets [1]. The findings reveal a significant supply chain security risk, with leaked secrets including GitHub personal access tokens, Docker Hub credentials, cloud service keys, and API keys for major AI providers [1].
The most critical discoveries include 349 valid GitHub tokens, of which 223 possessed full repository write permissions and 130 could modify CI workflows [1]. Additionally, researchers identified 318 Docker Hub tokens capable of pushing images, 8,557 valid Google Cloud service account keys, and access to 51.7 terabytes of protected AWS S3 storage buckets [1]. The leaked credentials also encompassed 8,594 verified database logins and 11,496 valid API keys covering OpenAI, Azure OpenAI, Anthropic, and Gemini services, including 742 OpenAI keys and 26 Anthropic keys [1]. Based on minimum consumption thresholds, the annual inference cost exposure amounts to at least $920,000 [1].
The scale of compromise extends beyond individual credentials. Forty-four percent of the leaked keys appeared across multiple datasets, with 19,380 credentials found in ten or more datasets; one Infura key appeared in 1,131 datasets alone [1]. The scan also surfaced 393 gigabytes of personal information covering approximately 3.7 percent of the global population [1]. The research team has collaborated with HuggingFace to responsibly disclose these findings [1].