Service Limitations & Important Notes
PII Firewall is a powerful AI security assistant. Like all technology tools, please read and understand the following before use.
Japanese Name Detection: Partial Space-Free Support (In Progress)
We have been progressively improving Japanese name detection. Here is the current status:
| Input | Detected | Method |
|---|---|---|
田中 太郎 (space-separated) | ✅ Detected | Space-separated (original method) |
田中太郎 (no space) | ✅ Detected | Surname dictionary matching (~400 surnames) |
田中さん / 長谷川様 | ✅ Detected | Honorific context detection |
林さん / 岡部長 | ✅ Detected | Single-char surname + honorific/title |
John Smith (English) | ✅ Detected | Space-separated |
| Rare surnames not in dictionary | ⚠️ Not detected | Use custom dictionary as workaround |
How it works:
The engine combines a dictionary of approximately 400 common Japanese surnames with honorific and title context patterns (e.g., さん, 様, 部長, 社長). Surnames in the dictionary can now be detected even without spaces between family and given name.
Workaround for names not in the dictionary:
- Register specific names as custom PII using the
extraTermsoption - Visually review masked output before sending to AI
Planned improvement: Continued expansion of the surname dictionary and further improvements to context-based analysis
🔬 NER Extension Option (Python SDK) — Enhanced Name Detection
For industries that require detection of rare surnames, katakana-format names, and furigana (reading) data, the Python SDK offers an optional GiNZa NER extension.
| Pattern | Standard (all platforms) | NER Extension (Python SDK) |
|---|---|---|
Yamada Taro (space-separated, dictionary name) | ✅ | ✅ |
ヤマダ タロウ (katakana) | ❌ | ✅ |
勅使河原 太郎 (rare kanji surname) | ❌ | ✅ |
ヤマダ様 (katakana + honorific) | ❌ | ✅ |
やまだ たろう (hiragana only) | ❌ | ⚠️ ja_ginza model limitation |
Recommended for these industries and use cases:
| Industry | Use Case |
|---|---|
| 🏥 Healthcare | Katakana patient names, non-Japanese patient records |
| 👥 HR / People Ops | Employee directories, payroll systems with furigana fields |
| ⚖️ Legal / LegalTech | Contract party names in varied formats |
| 🏦 Finance / Securities | Customer management systems with katakana name fields |
| 🏛️ Government / Public Sector | Katakana names of foreign residents (My Number integration) |
Installation:
pip install pii-firewall[nlp]Usage:
from pii_firewall import PIIFirewall
fw = PIIFirewall(ner_model="ja_ginza")
result = fw.mask_pii("Patient: ヤマダ タロウ, Visit date: 2026-05-01")
# → "Patient: [NAME], Visit date: 2026-05-01"⚠️ Platform availability
The NER extension is available in the Python SDK only. The Node.js SDK, REST API, MCP Server, and Chrome Extension use standard detection (dictionary + pattern matching). If you need equivalent accuracy in non-Python environments, please contact us.
PII Detection is an Assistive Feature
PII Firewall's detection engine combines statistical pattern matching with rule-based techniques. While detection targets and sensitivity are configurable, it is technically not possible to guarantee 100% coverage of all personal information.
What this means for you:
- We recommend visually reviewing masked prompts before submission, especially for sensitive use cases
- For high-stakes data (medical records, financial information, etc.), consider customizing detection rules
- Final responsibility for information management rests with the user
Masking & Restoration Accuracy
Masking and token restoration operate at high accuracy, but 100% completeness cannot be guaranteed due to technical constraints.
What this means for you:
- Review AI responses after token restoration for high-sensitivity content
- Restoration accuracy may decrease for very long or complex texts
Injection Detection is Pattern-Based
The prompt injection detection engine is based on known attack patterns collected and classified by our research team. Attack techniques evolve daily — we cannot guarantee detection of all attacks, including novel or unknown methods.
Our commitment:
- We continuously monitor the security research community for emerging threats
- New attack patterns are added to the detection engine as they are identified
- If you discover an attack technique we're missing, please let us know
Evolving Together with AI Technology
AI technology is advancing rapidly, and new use cases, attack vectors, and privacy risks emerge constantly. PII Firewall is committed to evolving alongside these changes.
- We regularly release updates covering new attack vectors and PII types
- Feature requests and bug reports are welcome at support@piifirewall.com or via GitHub Issues
- User feedback is what drives our improvements
Legal Disclaimer
For the full disclaimer and limitation of liability, please refer to Terms of Service, Section 6. By using the Service, you agree to these terms.