SafePrompt-Bench

1,800 synthetic documents of the kind people paste into AI chats — .env files, logs, JSON, CSV, curl commands, code, SQL, stack traces, e-mails and chat messages — with annotated personal data and secrets, and many look-alikes that are not sensitive.

Built by Since AI to evaluate SafePrompt, a local browser extension that checks prompts before they are sent. Generator and scorer: github.com/sinceaihq/safeprompt.

All data is synthetic. No real person, company, account or credential appears. Secret values are random strings in documented token formats with EXAMPLE or FAKE in them; they are not credentials.

Why

Existing PII benchmarks contain prose. Prompts often contain structured text: values sit in table columns and under JSON keys rather than after a label word, among UUIDs, hashes, timestamps, order numbers and barcodes. A detector that only understands "label: value" in sentences misses much of this — a direct port of EuroPII Guard found 42 % of national identifiers on the dev split.

Contents

splits	dev 300 documents (seed 4242) · test 1,500 documents (seed 20261020)
formats	env, log, json, csv (, ; or tab), curl, code (Python/JavaScript), sql, trace, email (en, fi, de, fr, sv, nl, es, it, pl), chat (values in running text without labels), business (no personal data), invalid_ids (no personal data)
labels (test counts)	EMAIL 1,432 · IBAN 1,064 · NATIONAL_ID 1,064 · PERSON 1,064 · PHONE 879 · SECRET 831 · IP_ADDRESS 189 · VAT_ID 71 · PAYMENT_CARD 67
national IDs	18 schemes: FI, SE, NO, DK, DE, FR, NL, BE, IT, ES, PL, EE, CZ, RO, HR, IE, AT, PT
secrets	AWS access key IDs, GitHub classic and fine-grained tokens, GitLab tokens, Google API keys, OpenAI and Anthropic keys, Hugging Face and npm tokens, JSON Web Tokens, passwords (in URLs and assignments)
hard negatives	UUIDs, SHA-1/SHA-256, base64, Unix timestamps, versions, order and invoice numbers, EAN codes, tracking codes, amounts, coordinates, IBANs and cards with wrong check digits, identifiers with a wrong check digit next to their label, placeholders such as <YOUR_API_KEY> and ${DB_PASSWORD}

Format

One JSON object per line:

{"id": "test-00000", "format": "sql", "text": "Why does this insert fail …", "spans": [{"start": 146, "end": 163, "label": "PERSON", "kind": ""}, {"start": 200, "end": 211, "label": "NATIONAL_ID", "kind": "fi_hetu"}]}

start/end are Unicode code point offsets into text (Python string indices). kind gives the identifier scheme (fi_hetu, pl_pesel, …) or secret type where relevant. Documents of the formats business and invalid_ids (and some others) have no spans.

from datasets import load_dataset
ds = load_dataset("sinceaihq/safeprompt-bench", split="test")

How the values were made

• National identifiers are valid under their official check-digit rules (checked with python-stdnum). Those that encode a birth date use real-world forms (e.g. the Finnish - century sign, 12-digit Swedish numbers) with birth years 1900–1905, verified by decoding them, so they look real without plausibly belonging to a living person. Identifiers without a birth date are random checksum-valid numbers not linked to anyone.
• Names are each country's conventional placeholder names (Matti Meikäläinen, Erika Mustermann, Jan Kowalski …).
• E-mail addresses use reserved domains (example.com, .example); IP addresses documentation (RFC 5737/3849) or private ranges; payment cards public test numbers.
• Phone numbers are libphonenumber example numbers with randomised last four digits, and IBANs random account numbers with valid check digits. They may coincide with real numbers but are not linked to any person.
• Stripe and Slack token formats are not included because code hosting platforms block repositories containing them even when they are fake.

Results (test split)

Personal data only (SECRET excluded), default settings, span overlap with the same label:

system	precision	recall	F1	clean documents flagged
SafePrompt 0.1.0	99.7	97.0	98.3	1 of 299
EuroPII Guard 0.1.2	93.7	83.9	88.6	79 of 299
Microsoft Presidio + EuroPII Guard	81.7	91.2	86.2	194 of 299
Microsoft Presidio (default, English)	80.8	70.5	75.3	177 of 299

SafePrompt's label handling was developed on the dev split; the others were not tuned on this data. See the full report for per-type and per-format results and the scoring rules. Secrets are not compared there: the generator uses the same documented formats SafePrompt detects, so such a comparison would be circular.

Limitations

• Synthetic and template-based: dev and test share templates; real prompts are more varied and results on them will be lower.
• One perspective on what counts as sensitive (for example, private IP addresses are labelled).
• PERSON spans are annotated but names appear in a few fixed forms; they are not suitable for training a name detector.
• Formats and languages are limited to those listed.

Licence and citation

CC BY 4.0. Generator code: Apache-2.0.

@misc{sinceai2026safepromptbench,
  title  = {SafePrompt-Bench: personal data and secrets in prompt-shaped text},
  author = {{Since AI}},
  year   = {2026},
  url    = {https://huggingface.co/datasets/sinceaihq/safeprompt-bench}
}