File size: 3,525 Bytes
5b9f6c9 14796cd 963f0b3 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 |
---
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.**
|