Build secure, reliable, and long-term AI systems.

Focused on safety, reasoning, and developer tooling.

We build offensive testing frameworks and deterministic defenses for engineering teams who value correctness over compliance. Neuralchemy focuses on measurable safety and rigorous stress-testing of AI models and software systems.

Get Started in 60 Seconds

Choose your path: Offensive Testing or Runtime Defense.

bash 
# Install from source
git clone https://github.com/neuralchemy/promptxploit
cd promptxploit
pip install -e .
python targets/http_api_target.py 
from promptxploit import HTTPTarget

# Configure your target API
target = HTTPTarget(
    url="https://your-api.com/chat",
    headers={"Authorization": "Bearer YOUR_TOKEN"},
    payload_template={"message": "{PAYLOAD}"},
    response_field="response"
)

# Run adaptive scan
target.scan(vectors=["injection", "jailbreak"])
python 
from huggingface_hub import hf_hub_download
import joblib

# Download 100% accuracy model
repo = "neuralchemy/prompt-injection-detector-ml-models"
vectorizer = joblib.load(hf_hub_download(repo, "tfidf_vectorizer_expanded.pkl"))
model = joblib.load(hf_hub_download(repo, "random_forest_expanded.pkl"))

def is_safe(text):
    return not model.predict(vectorizer.transform([text]))[0]

Products

PromptXploit

Open Source

Professional LLM Penetration Testing Framework

Comprehensive security testing framework with 147 pre-built attack vectors. Features adaptive AI-powered modes to discover vulnerabilities in AI apps before deployment.

View on GitHub

PromptShield

100% Accuracy v1 Available

Universal AI Security Framework

Production-ready defense protecting LLM apps from adversarial attacks. Backed by a comprehensive dataset of 10,674 samples.

View on GitHub

ReconRelate AI

Platform

Automated OSINT Relationship Discovery

Turns disconnected OSINT data into actionable threat intelligence. Automatically discovers and visualizes hidden relationships between domains.

Register Interest

Safety Regression

Infrastructure

Metric-Based Safety Regression Testing

Treats safety as an engineering metric. Runs adversarial scenarios against new logic to detect safety degradation ("Safety Diffs") before deployment.

Register Interest

Battle-Tested Performance

Production-grade security with zero compromise on speed.

10,674
Dataset Samples

Curated real-world attacks & synthetic variations

<1ms
Latency

Inference using optimized sklearn models

100%
Detection Rate

Zero false positives/negatives on test set

Open
Source

MIT Licensed & Free for commercial use

Philosophy

Security is an engineering discipline, not a checklist.

Adversarial Reality

Systems are only as secure as their ability to withstand active adaptation. We test against verifyable threats.

Measurement Over Vibes

Security claims must be verifiable and reproducible. If you can't measure it, you can't secure it.

Engineering First

Tools should integrate natively into developer workflows, not impede them. Frictionless security is effective security.

✓ Email copied to clipboard