1. Overview
- Original Report Title: GitHub Repos Exposed 543,699 Credentials. Nobody Revoked Them.
- Publisher: Truffle Security
- Publication Date: 2026-09-29
- Source: Original GitHub credential study
- Related Sources: BleepingComputer, GitHub secret-scanning patterns, Secrets Revocation Program, BleepingComputer: 1.4 million live secrets
- Severity: High
- Basis for Severity: The research measures substantial verified credential exposure, long estimated ages based on file timestamps, and gaps in default protection coverage. The November update adds issuer-coordinated revocation.
2. Executive Summary
Truffle Security's original study analyzed 543,699 credentials from public GitHub repositories that authenticated when tested in July 2026 and could be associated with usable file timestamps.
In a November update, it reported a cumulative 1.4 million exposed secrets verified since 2022 across GitHub, GitLab, container registries, package repositories, datasets, and applications. It launched a Secrets Revocation Program to work directly with issuers.
The cumulative figure is not a count of secrets all still valid today, nor a count already revoked.
The reports also distinguish notification from revocation. In one notified GitLab/Bitbucket population, 81% remained valid six months later. A separate Google Cloud comparison found 72% valid in a pre-policy group and 11% in a post-policy group. Those are different populations and study designs.
3. Scope and Methodology
The original analysis used The Stack v3: 224,553,295 repositories and 58,467,468,698 file entries from a crawl ending August 7, 2025.
It covered default-branch snapshots, not full commit history or deleted branches.
Evaluation Design
- Credential validity was tested on July 27–28, 2026.
- Credentials were deduplicated by value.
- The earliest associated file timestamp was used as a proxy for exposure timing.
- Google API keys were counted as valid only when they authenticated to Gemini; other Google services were not tested.
- The Google API-key family was excluded from both numerator and denominator in survival-rate calculations because that validation scope was incomplete.
The later multi-source revocation program is broader than this GitHub snapshot study.
4. Key Findings
- The primary analysis set contained 543,699 verified credentials with usable timestamps.
- The estimated interval from file timestamp to verification had a median of 784 days, a 90th percentile of 6.3 years, and a maximum of 16.1 years.
- Of the analysis set, 199,843 credentials, or 36.8%, had file timestamps on or after February 29, 2024, when GitHub enabled Push Protection by default.
- Separately, 51.8% of the entire analysis set fell outside default blocking coverage under the study's classification.
- Comparing the 12-month periods on either side of March 2024, median monthly credential density per million files declined by 53% for protected families and 7% for unprotected families.
The density comparison is observational. It is not the proportion of individual pushes blocked.
Provider-Specific Verification Results
| Credential Type | Candidates Tested | Verified as Valid | Calculated Rate and Limit |
|---|---|---|---|
| npm tokens | 101,886 | 1 | Approximately 0.001%; failed verification does not establish revocation as the sole cause. |
| GitHub tokens | 73,048 | 260 | Approximately 0.36%; this does not isolate the causal effect of automatic revocation. |
| Google Cloud service-account credentials | 126,963 | 69,041 | Approximately 54.38%; the denominator is this candidate population, not the 543,699 valid-credential analysis set. |
5. Interpretation and Limitations
- Snapshot Limits: The corpus excludes commit history and deleted branches. File timestamps are proxies, not verified first-exposure dates.
- Continuity Limits: A long interval between a file timestamp and a successful test does not directly establish uninterrupted validity throughout that interval.
- MongoDB: The detector reports connection strings after successful connection, so MongoDB was excluded from survival-rate and exposure-density calculations.
- Private Keys: A private-key value alone cannot establish its validity against an authorized counterpart; these keys were excluded from survival calculations.
- Failed Verification: Failure can reflect issuer-side deletion as well as revocation.
- Misuse: The study did not measure whether third parties had abused the credentials.
- The 51.8% Aggregate: The discussion mentions private keys while the methodology excludes them from validity testing or survival analysis. Their precise treatment in the aggregate requires clarification against the underlying data.
5.1 November 2026 Update: Issuer-Coordinated Revocation
Program Scope
Truffle reported a cumulative 1.4 million secrets that passed validity checks since 2022 across sources including GitHub, GitLab, Docker Hub, npm, PyPI, Hugging Face, Android applications, and Firefox add-ons.
The company reported scanning 6 PB in October 2026 and more than 20 PB during the year. Under the program, a key is verified before it is delivered to its issuer.
Historical cumulative verification, validity at delivery, and completed revocation are separate measures.
Two Revocation Models
- Direct Revocation: Truffle calls an issuer-provided revocation endpoint; the issuer handles customer notification.
- Webhook Delivery: The issuer receives verified keys and their discovery locations, then revokes them within 14 days or explains why it cannot.
At launch, 16 issuers had joined and 43,699 verified keys were being held for issuers that had not joined. These figures are not a completed-revocation rate.
Different Observations Must Remain Separate
| Observation | Population and Timing | Result | Interpretation Limit |
|---|---|---|---|
| Notification follow-up | 21,981 keys reported from GitLab and Bitbucket; retested six months later | 81% remained valid | A follow-up of notified keys, not a randomized comparison of controls |
| Push Protection comparison | Monthly credential density in the 12 months before and after March 2024 | Protected families down 53%; unprotected families down 7% | Not a six-month survival rate or individual push-blocking rate |
| Google Cloud policy comparison | The Stack v3 groups split around June 16, 2024; validity checked in July 2026 | 72% valid before the policy; 11% after | Not equal six-month follow-up and not a randomized estimate isolating policy effects |
| Revocation-program launch | Participating issuers and keys awaiting non-participating issuers | 16 issuers; 43,699 pending keys | Does not establish how many keys were revoked or how quickly |
AI-Related Findings
Truffle reported 2,655 working keys in AI coding-tool configuration files since January 2025.
By May 2026, AI-agent-signed commits accounted for 6.8% of commits that leaked a working key in the reported analysis. The denominator is leaking commits, not all commits or all exposed keys. Signature classification and sampling bias require further examination.
6. Facts, Inference, and Hypothesis
Facts
- Verification identified 544,069 distinct valid credentials. Of these, 543,699 had the corpus mapping and usable timestamps required for the main analysis.
- The analysis set's estimated median age was 784 days, based on file timestamps rather than directly observed first exposure.
- The 51.8% coverage figure uses the entire analysis set, not only the 199,843 credentials dated after default Push Protection began.
- The density comparison used six protected and six unprotected credential families already present in 2021 and excluded MongoDB.
- The provider-specific candidate populations differ from the valid-credential analysis set.
- The November update reported 1.4 million cumulatively verified secrets since 2022 and 6 PB scanned during October.
- The 81% notification result and the 72% versus 11% Google Cloud comparison have different populations and timing.
Inference
- Preventing a push and revoking an already exposed credential are separate controls.
- Failed verification cannot always be attributed to automated revocation.
- Before-and-after density comparisons cannot exclude concurrent changes such as short-lived credentials or workload identity.
- Issuer-coordinated revocation can shorten exposure, but verification, owner notification, emergency rotation, exceptions, and service recovery are needed to manage disruption.
Hypothesis
No additional hypotheses. Unconfirmed matters are listed in Section 7.
7. Open Questions and Further Investigation
- What proportion of the original credentials were obtained or misused by third parties?
- Which owners were notified, and which rotations were completed?
- What remains valid outside default branches, deleted history, or exposures after the snapshot?
- How many of the cumulative 1.4 million secrets have actually been revoked?
- What are the time-to-revoke distribution, incorrect-revocation count, service-disruption impact, and webhook deadline compliance?
- How were AI-agent-signed commits identified, and how do repository or issuer biases affect the 6.8% figure?
8. Implications for Defenders
Track exposed credentials from detection through ownership, issuer receipt, misuse investigation, revocation, replacement, and confirmation that the old credential no longer works.
Push blocking is valuable but does not remove historical exposure or cover every format. Scan history and separately assess connection strings, private keys, and Google API or Gemini keys according to the relevant protection scope.
Revoke exposed credentials before relying on repository cleanup. Where applicable, review and revoke sessions separately; replacing a credential is not proof that every derived session has ended.
Deployment and Operational Guidance
- Revoke committed credentials promptly, then clean up repository history.
- Add coverage for generic secret patterns and database connection strings.
- Prefer short-lived credentials and workload identity where supported.
- Standardize verified-key intake, emergency revocation, owner notification, and service recovery.
- For the program's webhook model, operationalize the 14-day revocation-or-explanation requirement. It is not a universal SLA for every issuer.
Required Evidence
Correlate alerts, repository and commit references, credential fingerprints, owners, verification times, issuer receipt, revocation responses, replacement issuance, authentication history, and confirmation that old credentials are unusable.
Keep secret values out of shared incident records.
9. Summary by Audience
- SOC: Connect code-exposure alerts with identity and cloud audit records, and follow cases through issuer receipt and confirmed revocation.
- Administrators: Scan history, identify coverage gaps and long-lived credentials, and move toward short-lived credentials and workload identity.
- Users: Do not store credentials directly in code. If a secret is committed, deletion alone is insufficient; report it and arrange immediate revocation and replacement. 最新の応答
GPT-6.1 Sol
中
なし
最小
軽
中
高
極高
マックス
ウルトラ
パーシステント
出力
出力を作成
ソース
ウェブ検索
GitHub
Top comments (0)