Skip to content

Response, New York. Noah Park: Check Point Patches Actively Exploited Management Server Path Traversal. Detection, Dallas. Ava Okello: Report Finds Nearly Half of Deployed Detections Need Attention Before SOCs Can Trust Them.

SOCtember

Always First. Fast SOC News.

MongoDB-published bar chart of unique verified-live credentials found per vendor in public Gitee repositories, led by Alibaba Cloud, YouTube API and DeepSeek.
Image: Ben Zimmermann / MongoDB (fair use). Unique verified-live credentials per vendor across 7.25 million Gitee repositories.

Tools

A Scan of 7.25 Million Gitee Repositories Found 26,400 Live Credentials, and TruffleHog and Kingfisher Each Caught Only About Half, Agreeing on Just 1,001

Independent researcher Ben Zimmermann ran both open-source scanners on every repository and kept only secrets that still authenticated. Alibaba Cloud keys led at 9,473, an organization-admin credential for an unnamed enterprise AI platform turned up, and revocation of the reported keys had not been confirmed when he published.

James Whitford, ToolsAustin4 min read

AUSTIN - An independent researcher who cloned and scanned 7,256,531 public repositories on Gitee, China's largest homegrown Git hosting platform, found 26,400 unique credentials that still authenticated against the services that issued them, and the two open-source secret scanners he ran side by side each caught only about half of them.

The findings, published Oct. 28 by Ben Zimmermann as a guest post on MongoDB's blog and developed with MongoDB's Kingfisher team, include what he described as "an active organization-admin credential for a major enterprise AI platform" and thousands of live cloud provider access keys. He did not name the platform.

Two scanners, little overlap

Zimmermann ran TruffleHog and MongoDB's Kingfisher concurrently against a bare clone of every repository, with each tool set to keep only secrets that verified against the vendor's API. TruffleHog reported 12,737 unique live credentials and Kingfisher 14,664. Only 1,001 were caught by both. Deduplicated by raw secret value, the total was 26,400.

Diagram of the scan pipeline: a Modal worker bare-clones each repository and feeds TruffleHog and Kingfisher concurrently, writing findings to a shared volume.
Image: Ben Zimmermann / MongoDB (fair use).The dual-scanner pipeline, run at 350 concurrent workers.

The gap came down to detector coverage. TruffleHog dominated Alibaba Cloud key detection, 9,428 to 993. Kingfisher dominated DeepSeek, 2,940 to 230; OpenAI, 1,198 to 150; Hugging Face, 154 to 1; and Zhipu's BigModel platform, 1,489 to zero, because TruffleHog has no Zhipu detector. "Either scanner alone would have caught only about half of these credentials," Zimmermann wrote.

For a team that runs one scanner in CI or across its own repositories, the lesson is narrow and practical: a clean result means only that the scanner's detectors found nothing. If the organization uses cloud or AI providers a tool does not cover, those keys pass through unseen. Checking the detector list against the providers actually in use, or running a second engine, closes that gap.

What leaked, and when

Alibaba Cloud AccessKey pairs were the most common type, at 9,473 unique credentials, followed by 3,805 YouTube API keys and 3,170 DeepSeek keys, according to the post's vendor chart. A long tail included LangSmith, OpenWeather, Algolia, PostHog, OpenRouter, MySQL, Stripe and Slack.

Line chart of unique verified-live credentials by commit year from 2008 to 2025, peaking at 12,242 in 2025.
Image: Ben Zimmermann / MongoDB (fair use).Live credentials by commit year.

Most of the live credentials were committed in the last two years; 2025 alone accounts for 12,242, nearly half the dataset. Because both scanners read full commit history, a key a developer later deleted from the code still counted if it was in an earlier commit and still worked.

Bar chart of unique verified-live credentials by committer email domain, including huawei.com, microsoft.com and bytedance.com.
Image: Ben Zimmermann / MongoDB (fair use).Live credentials by committer email domain.

Most commits came from personal mail services such as qq.com, 163.com and gmail.com, so Zimmermann said per-company attribution was impossible at scale. Among corporate domains, his chart shows huawei.com with 32 credentials and microsoft.com and bytedance.com with 22 each, and he said smaller footprints appeared from Oracle, Apple, Meta, Anthropic and others.

Density and method

Gitee's rate worked out to 3.64 verified-live credentials per 1,000 repositories, compared with 3.11 for about 5.6 million GitLab repositories and 2.39 for about 2.6 million Bitbucket repositories in earlier Truffle Security scans. Zimmermann cautioned that those scans used TruffleHog alone; counting only his TruffleHog results, Gitee's rate falls to 1.76.

Enumeration was the hard part, because Gitee's public API is limited. He combined the one official REST endpoint that returned useful data at scale, the Explore feed, per-user repository and star pages, and the internal search widget behind Gitee's site search, splitting capped queries by language, license and prefix. The scan ran as 50 batches of about 150,000 repositories at 350 concurrent containers on Modal for roughly a month and cost $3,097.64, covered by a $5,000 Modal compute credit.

Revocation not confirmed

Zimmermann emailed the vendors with the most exposed credentials, giving each the full list with repository metadata, but did not contact individual committers. At the time of writing, he said, revocation of the disclosed credentials had not been confirmed. Removing a key from code or rewriting history does not revoke it at the provider, so any key that has ever sat in a public Gitee repository should be treated as exposed and rotated.

Sources:


James Whitford covers SOC tooling, SIEM, SOAR, and detection platforms for SOCtember from Austin.

Related stories

Tools desk