---
title: README
emoji: 🌖
colorFrom: gray
colorTo: red
sdk: static
pinned: false
---

# 🌖 Durinn — AI Security

Durinn builds **AI security infrastructure** for high-assurance and regulated environments.  
Our work focuses on **calibration**, **dataset poisoning detection**, and  
**neuro-symbolic vulnerability analysis** for safer, more predictable agents.

We contribute research datasets, calibration tools, and security-focused evaluation  
pipelines designed for GxP, healthcare, and enterprise LLM deployments.

---

## 🧪 Research Focus

Our work spans:

- Calibration of high-stakes LLM security classifiers  
- Prompt-injection detection  
- Dataset poisoning defense  
- Neuro-symbolic vulnerability scoring  
- Evaluation and benchmarking for regulated AI systems  

Our Hacktoberfest-derived dataset supports real-world model calibration and  
has demonstrated meaningful improvements when applied to production-grade PI classifiers.

---

## 🧭 Agent Safety, Guardrails & Calibration

Durinn calibrates state-of-the-art prompt-injection classifiers, including models  
widely deployed in production security pipelines.  

Calibration improves:

- Detection of subtle prompt injections  
- Threshold placement (better true-positive recovery)  
- Agent stability and predictability  
- Decision-level robustness for regulated environments  

These calibrated guardrails can be deployed in:

- **Internal inference pipelines** as an agent heartbeat  
- **AIDR / SOC / cloud platforms** enhancing their LLM input-security layers  


---


## 🧬 Dataset Poisoning & Model-Integrity Defense

Our work includes:

- Poisoning detection in training and inference datasets  
- Checkpoint tampering & backdoor forensics  
- Model-integrity drift analysis  
- Provenance and chain-of-custody guidance for regulated AI stacks  

We emphasize **verifiable integrity** for teams who cannot rely on opaque model behavior.


---

## 🔐 Neuro-Symbolic Vulnerability Detection

Durinn develops hybrid detection approaches that combine:

- Symbolic signals from program analysis  
- LLM reasoning  
- Safety-critic scoring  
- Calibrated confidence thresholds  

This architecture improves reliability without altering underlying model weights.


---
## 📚 Key Repositories

- **`durinn-calibration`** — Tools and experiments for calibrating security-critical classifiers, including prompt-injection detectors and safety-critic models. Contains evaluation scripts, threshold-optimization utilities, and datasets for benchmarking calibrated decisions in regulated AI environments.
- **`durinn-sandbox`** — A high-assurance execution environment for analyzing model behavior, running controlled adversarial tests, and validating agent outputs. Provides reproducible sandboxes for measuring failure modes, safety drift, and poisoning-related anomalies.
- **`durinn-agent-infrastructure`** — Shared infrastructure components for constructing and evaluating secure AI agents. Includes model wrappers, risk-scoring pipelines, input-validation hooks, telemetry collection, and integration utilities for enterprise inference stacks.
- **`durinn-ai-code-remediation`** — Research agent for neuro-symbolic vulnerability detection and compliant secure-rewrite workflows. Designed for GxP and regulated industries requiring traceability, safety justification, and audit-aligned remediation artifacts.

---

**Durinn — Secure, calibrated, and trustworthy AI for environments where accuracy and integrity matter.**