AI & ML interests
None defined yet.
LangProtect š”ļø
Securing LLM applications, one layer at a time.
LangProtect builds open and commercial-grade AI security models that help developers detect, block, and monitor threats in LLM-powered applications ā starting with prompt injection detection and extending into domain-specific NER for regulated industries like fintech.
š Website: www.langprotect.com
šÆ What We Do
As LLMs get embedded deeper into products ā chatbots, agents, RAG pipelines, autonomous workflows ā they become exposed to a new class of attacks: prompt injection, jailbreaks, and data leakage. Traditional security tools weren't built for this.
LangProtect focuses on:
- Prompt Injection Detection ā lightweight, fast classifiers that flag adversarial or manipulative inputs before they reach your LLM.
- Named Entity Recognition (NER) for Fintech ā models fine-tuned to identify financial entities (transactions, account details, PII, instruments, etc.) for compliance, redaction, and data governance use cases.
- Applied AI Security Research ā iterating on datasets and models to keep pace with evolving attack patterns.
š§ Models
| Model Type | Task | Description |
|---|---|---|
| Prompt Injection Classifier (latest) | Text Classification | Detects adversarial/malicious prompts aimed at manipulating LLM behavior. |
| Prompt Injection Classifier (earlier generation) | Text Classification | Previous-generation baseline for injection/jailbreak detection. |
| Fintech NER Model (latest) | Token Classification (NER) | Transformer-based model fine-tuned to extract financial entities from text. |
| Fintech NER Model (prior versions) | Token Classification (NER) | Earlier iterations used during development and benchmarking. |
| Fintech NER Baseline | Token Classification (NER) | Initial baseline model for financial entity extraction. |
| Toxicity Detection Model | Text Classification | Flags toxic, abusive, or harmful language in user inputs and LLM outputs. |
| Bias Detection Model | Text Classification | Identifies biased or discriminatory language patterns in text. |
| PII Detection Model | Token Classification (NER) | Detects personally identifiable information (names, emails, phone numbers, IDs, etc.) for privacy and compliance workflows. |
| Secrets Detection Model | Token Classification (NER) | Detects leaked credentials, API keys, tokens, and other secrets in text for IT/security compliance. |
ā Our flagship model: the latest Prompt Injection Classifier is our best-performing release to date, built on continuous benchmarking against evolving attack patterns ā and it's the model we recommend as a first line of defense for any LLM application.
š Note: All models are currently private while under active development. Public model cards with usage examples, benchmarks, and licensing details will be published as models mature.
š Datasets
| Dataset Type | Description |
|---|---|
| Prompt Injection Dataset | Labeled examples of benign vs. adversarial/injection prompts, used to train our classifiers. |
| Fintech NER Dataset (latest) | Annotated fintech text data used to train our current-generation NER model. |
| Fintech NER Dataset (prior versions) | Earlier annotated datasets used during iterative model development. |
| Fintech NER Baseline Dataset | Initial fintech entity dataset used for our baseline model. |
| Toxicity Detection Dataset | Labeled examples of toxic vs. non-toxic text for training our toxicity classifier. |
| Bias Detection Dataset | Annotated examples of biased/discriminatory vs. neutral language. |
| PII Detection Dataset | Annotated text containing personally identifiable information for entity-level detection training. |
| Secrets Detection Dataset | Annotated examples of leaked credentials, keys, and tokens used for IT compliance model training. |
š Note: All datasets are currently private. Curated public releases (with documentation and licensing) are planned for the future.
š Why LangProtect
- Purpose-built for LLM security ā not a repurposed generic classifier, but models trained specifically on injection/jailbreak patterns.
- Domain-specific accuracy ā fintech NER models tuned for the entity types and language actually found in financial text.
- Iterative & versioned ā we continuously retrain and release improved versions (v1.7 ā v2.0) as new attack patterns and data emerge.
- Lightweight & production-ready ā designed to run as a fast pre-filter layer in front of your LLM, not a heavyweight bottleneck.
š ļø Use Cases
- Pre-filtering user inputs to LLM apps/agents for injection attempts
- Auditing/red-teaming LLM pipelines against adversarial prompts
- Automated redaction/compliance screening of financial documents
- Building safer RAG and customer-support chatbot pipelines in regulated industries
- Moderating LLM inputs/outputs for toxic or abusive content
- Screening text for biased or discriminatory language
- Detecting and redacting PII for privacy compliance (GDPR, CCPA, etc.)
- Scanning code, logs, and text for leaked secrets/credentials as part of IT security compliance
š¬ Get in Touch
- š Website: langprotect.com
- āļø Email: team@langprotect.com
- š¤ Follow us here on Hugging Face for model releases and dataset updates
If you're using LangProtect models in production or research, we'd love to hear about it ā feel free to open a discussion on any of our model/dataset repos.