password-detector

The main idea: keep it small and fast. A model that runs anywhere on a CPU, with no GPU and no heavy dependencies.

A tiny model that finds passwords, API keys, tokens and private keys in any text: code, configs, logs, emails, chats and documents. Byte-level CNN, no tokenizer, runs on one CPU core with numpy + onnxruntime. From 97 KB (tiny) to 0.7 MB (recommended cascade).

Quick start

pip install huggingface_hub numpy "onnxruntime>=1.17"
huggingface-cli download bardsai/password-detector --local-dir password-detector && cd password-detector

from password_detector.detector import SecretDetector

det = SecretDetector("onnx/default/model_int8.onnx,onnx/large/model_int8.onnx",
                     prefilter="onnx/tiny/model_int8.onnx")
det.scan("Cześć, hasło do VPN to kawa2024!biuro")
# [{'value': 'kawa2024!biuro', 'score': 1.0, 'type': 'password', 'start': 23, 'end': 37}]
det.scan("password = os.environ['DB_PASSWORD']")
# []

CLI (values masked unless --show; exit code 1 when anything is found):

python -m password_detector --model onnx/tiny/model_int8.onnx path/to/files

Types: password, api_key, token, private_key, generic. Default threshold 0.5.

Results

Offline, one CPU core, int8 models at threshold 0.5.

Model	Size	CredData F1	GitHub issues F1	Docs F1	Speed
cascade (tiny → default + large)	0.7 MB	0.89	0.85	0.94	138 KB/s
tiny (onnx/tiny)	97 KB	0.87	0.85	0.94	137 KB/s
default (onnx/default)	219 KB	0.88	0.87	0.93	66 KB/s
CredSweeper 1.19 *	12 MB	0.94	0.79	0.41	230 KB/s
detect-secrets 1.5	1.1 MB	0.57	0.48	0.32	155 KB/s
gitleaks 8	23 MB	0.44	0.69	0.25	958 KB/s

CredData: 64 held-out repos of Samsung CredData. GitHub issues: Ahmed et al. 2025. Docs: synthetic emails, chats and tickets in 5 languages. * CredSweeper was trained on all of CredData, so its CredData score is not held out.

Limitations

• Passwords in prose with unusual wording (short PINs, "access code X") can score below 0.5.
• Hex strings and UUIDs next to words like secret sometimes raise false alarms.
• Does not verify tokens with providers or detect secrets split across lines.

Training data: CredData training repos, synthetic code, configs and documents (randomly generated secrets) and real non-secret text. No benchmark data was used for training.