Detection reference · Credentials and secrets · Updated 2026-11-03
Hugging Face user access tokens (hf_)
Hugging Face tokens end up in notebooks and training scripts more often than anywhere else. A write token can push to your models and datasets.
At a glance
| Reported as | API key or token |
|---|---|
| Category | Credentials and secrets |
| Severity | critical: direct account or system compromise, as with keys, passwords and tokens |
| Replaced with | [SECRET] |
| Where it runs | The same engine runs in the Chrome extension (free checks, on your device), in the web check and the Windows app (on Redaxa's server; the text is not stored) and in public repository checks. |
| Engine rule | scanner.ts, line 205 |
What it looks like
hf_ followed by exactly 34 letters or digits.
Example
You paste
login(token="hf_RedaxaDocsSampleNotReal0RedaxaDocs")
Redaxa returns
login(token="[SECRET]")
| Reported as | API key or token |
|---|---|
| Severity | critical: direct account or system compromise, as with keys, passwords and tokens |
| Replaced with | [SECRET] |
| Probable service | Hugging Face |
| Credential type | User access token |
| Evidence | Hugging Face-specific prefix; account, role and validity are not verified. |
| Recommended action | Revoke or rotate if genuine |
All values on this page are synthetic or published specimens; none is a working credential or a real person's data.
In a repository check
The same example, committed to a public repository and run through the free repository check:
notebooks/train.ipynb: API key or token, listed for review (severity critical). Credential pattern matched. Validity is not tested; review and rotate if genuine.- Same value in
tests/fixtures/train.ipynb: Test or example value, counted but not listed. Real-looking keys in test files are still reported.
Values are always shown masked in repository reports. No key is ever tested against its provider.
What it does not flag
login(token="hf_xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx"): not flagged. A body of x is a placeholder.
If it already leaked
Invalidate the token in your Hugging Face settings (Access Tokens), create a fine-grained replacement, and review recent activity on your models and datasets.
What Redaxa adds to every finding of this kind: If genuine, revoke or rotate it at the issuing service, update the application, then remove the exposed value from code and review Git history. Deleting it from Git does not revoke access.
Official guidance: huggingface.co.
Step by step: pasted a key into ChatGPT, or committed one to GitHub.
Check before you paste
The free Chrome extension runs this check inside ChatGPT, Claude, Gemini and Perplexity before a message is sent, on your device and with no daily limit. On the web you get 5 free prompt checks a day with no account.
Add the Chrome extensionTry a free web checkWorried it is already in your code? Check a public GitHub repository free (3 checks a day).
Related
- GitHub tokens
- GitLab personal access tokens
- OpenAI API keys
- Anthropic API keys
- Stripe keys and webhook secrets
- Supabase keys
- Slack tokens and webhook URLs
- Google API keys
- AWS access keys
- npm access tokens
- PyPI API tokens
- SendGrid API keys
- JSON Web Tokens (JWT)
- Bearer tokens
- Passwords in URLs
- Secrets in configuration values
- Passwords written in text
- Private keys (PEM)
- I pasted an API key into ChatGPT: what to do
- Stop API keys leaking into AI prompts
- Check a public GitHub repository for leaked API keys
- Free GitHub secret scanner
- Everything Redaxa detects
Sources
- Redaxa engine: scanner.ts line 205, credential-context.ts line 39 (Hugging Face). Every output on this page was produced by that code when the page was built.
- Hugging Face Docs: User access tokens