26,400 Valid Cloud and AI Credentials Exposed in Public Gitee Repositories
Security researcher identifies thousands of active API keys leaked within Chinese code-hosting platform repositories.
1Massive Credential Exposure Identified
Researcher Ben Zimmermann discovered 26,400 active credentials exposed within 7.25 million public Gitee repositories. These keys, ranging from cloud services to AI platforms, remained functional during testing. The scan, which spanned approximately one month, highlights a critical security vulnerability where developers inadvertently push sensitive authentication tokens into public code-hosting environments.
2Leading Platforms at Risk
The exposed keys include 9,473 Alibaba Cloud AccessKey pairs, representing the largest group identified. Other significant findings include 3,805 YouTube API keys, 3,170 DeepSeek keys, and 1,348 OpenAI keys. Additionally, researchers identified an active organization-admin credential for a major enterprise AI platform, though the specific service provider remains unnamed in the report.
3Technical Scanning Methodology
Zimmermann utilized two specialized secret scanners, TruffleHog and Kingfisher, to audit full commit histories across all repositories. These tools validated credentials by testing them against their respective issuing services. The scanners identified 12,737 and 14,664 working keys respectively, with only 1,001 overlaps, demonstrating that relying on a single detection tool is insufficient.
4Persistence of Leaked Secrets
Deleting a key from a repository or rewriting commit history does not invalidate the credential. Once a key is pushed to a public repository, it must be explicitly revoked at the provider level to prevent unauthorized access. Developers often mistakenly assume that removing the code from the repository history effectively secures the compromised secret.
5Comparative Security Analysis
Gitee exhibits a rate of 3.64 working credentials per 1,000 repositories, surpassing previously reported rates for GitLab at 3.11 and Bitbucket at 2.39. However, direct comparisons remain difficult due to varying scanner coverage and service support. For instance, Kingfisher identified all Zhipu BigModel keys, whereas TruffleHog lacked the specific detector for that service.
6Temporal Distribution of Leaks
Analysis of commit dates reveals that 12,242 of the exposed credentials originated from 2025, accounting for nearly half of the total findings. Some of the leaked keys date back to the late 2000s. This persistence underscores the long-term risk of failing to rotate credentials that have ever been exposed in public code.
7Operational Impact of Compromise
Compromised keys allow unauthorized parties to utilize account quotas, potentially leading to data loss or unexpected financial charges. For services like YouTube, an exposed API key identifies a specific project and can be leveraged across all APIs accepting Google Cloud keys. Such access can result in interrupted services and significant security breaches.
8Strategic Security Recommendations
Organizations must treat any key ever present in a public repository as compromised and perform immediate revocation. Security teams should verify their scanning tools against the specific services they utilize to ensure comprehensive coverage. Proactive identity visibility and robust incident response protocols are essential to mitigate threats across human, machine, and AI identities.