What a Scan of 7.26 Million Gitee Repositories Found: 26,400 Working Credentials
A scan of 7.26 million public Gitee repositories found 26,400 live credentials, nearly 9,500 of them for Alibaba Cloud. Here is what that means for cloud and AI users.

An independent security researcher scanned 7,256,531 public repositories on Gitee, the Chinese code-hosting platform, and reported 26,400 unique credentials that still authenticated when tested. The write-up was published on MongoDB's blog on 28 October. The largest group was Alibaba Cloud AccessKey pairs, with 9,473 unique credentials.
For companies that run workloads on Alibaba Cloud or on global cloud and AI platforms, the findings show that a secret committed to a repository can still authenticate when it is tested. The keys in the dataset also cover AI model providers, payments and messaging services, so the exposure is not limited to one vendor.
How the Scan Reached Live Credentials
Gitee's public API was too restricted for a full crawl. The researcher combined the official user-search endpoint, the public Explore feed, per-user repository pages and an internal search widget that capped results at 200 per query. That produced a seed list of 1,704,759 usernames.
Repositories were cloned over SSH on serverless compute, in 50 batches of about 150,000 repositories each. The run took about a month, and clones were slower than on GitHub, partly because Gitee's servers are in China. The total cost was $3,097.64, covered by a $5,000 compute-credit sponsorship from Modal.
Each candidate secret was tested to see whether it still worked, and duplicates were removed by raw value. The 26,400 figure therefore counts only secrets that authenticated, not every match in the code.
Which Providers Showed Up Most Often
The write-up also mentions LangSmith, OpenRouter, OpenAI, Zhipu, HuggingFace, Stripe, Slack and MySQL, among others.
Most of the credentials were committed recently. Keys added in 2025 account for 12,242 of the unique credentials, nearly half the dataset.
Alibaba Cloud keys appeared in roughly one of every 766 repositories scanned. The write-up calls Alibaba Cloud AccessKey pairs the single most common leaked credential type.
Two Scanners, and Why Their Overlap Matters
The researcher ran TruffleHog and Kingfisher against each repository. TruffleHog found 12,737 unique credentials, Kingfisher found 14,664, and 1,001 were caught by both tools.
The tools differed in coverage. TruffleHog was stronger on Alibaba Cloud keys, while Kingfisher dominated DeepSeek, OpenAI, HuggingFace and Zhipu keys. TruffleHog has no detector for Zhipu.
The cross-platform comparison is less direct. The write-up lists 3.64 verified credentials per 1,000 Gitee repositories, against 3.11 for GitLab and 2.39 for Bitbucket. The GitLab and Bitbucket figures came from TruffleHog-only scans, and on TruffleHog alone Gitee comes to 1.76 per 1,000. The higher Gitee rate therefore depends partly on the second scanner.
Vendor Revocation and the Notification Gap
The researcher emailed the vendors with the highest exposure volume, sending a coordinated report with affected credentials and source-repository details. Individual committers were not notified; the write-up says per-developer notification was not pursued at this scale.
As of the write-up, revocation had not been confirmed for the disclosed credentials. The write-up describes vendor-side revocation as the more reliable path, because it does not depend on the person who committed the key.
Alibaba Cloud's own guidance says the company notifies customers when it detects publicly exposed keys and applies restrictive protection to them. It also says it cannot monitor the security status of every key pair, and that customers remain responsible for protecting their own keys under the shared responsibility model.
The write-up does not say whether any of the 26,400 keys have since been revoked.
“Vendor-side revocation is the more reliable path, since it doesn't require any action from the user who leaked the credential.”
Rotating an Alibaba Cloud AccessKey Step by Step
Alibaba Cloud allows up to two AccessKeys per RAM user, so a second key can be created before the original is removed. The procedure is to create a second key, move every application and system to it, disable the original, confirm everything works, and then delete the original.
Before disabling, check the last-used time for the original key in the RAM console. If an application fails after the disable, re-enable the original immediately and repeat the steps until the switch succeeds.
Keys in use for more than 90 days should be rotated, and the cycle repeats every 90 days or sooner if needed. The guide was last updated on 10 July 2026.
Leak Response and Secret Scanning Requirements
Alibaba's leak guidance starts with disabling or deleting the key and reviewing its activity in ActionTrail over the previous 90 days. Data-plane logs for services such as OSS and SLS should be queried with each service's own log feature. Reviewers should look for new users, new keys, attached policies, deleted resources and changes to bucket access.
Before disabling or deleting a key, the guidance says to check local machines, CI/CD pipelines, containers, scheduled jobs and third-party apps for overlooked callers. A deny policy can restrict high-risk actions while the key is still in use, and billing should be checked for unusual charges.
On the repository side, Alibaba recommends removing plaintext keys, storing them in environment variables or its key management service, restricting use with IP allowlists and scanning code hosted on third-party platforms. Keys in private repositories can still leak through forks, weak access controls, server compromise or application flaws.
For secret scanning, the Gitee dataset shows that no single tool was enough. TruffleHog found about 48% of the verified keys and Kingfisher about 56%. Together they account for all 26,400, so a company that scans with only one tool would leave gaps.



