v0.4.0 — 12 LLM Providers · 9 SDK Integrations

The AI Firewall
That Fights Back

Detection, deception, and intelligence for LLM security. Don't just block prompt injections — trap attackers with honeypots while generating threat intelligence.

$ pip install oubliette-shield
F1: 0.98 2ms ML Inference Low False Positive Rate Apache 2.0
Learn More View on GitHub
quickstart.py 
from oubliette_shield import Shield

# Initialize with any LLM backend
shield = Shield(
    provider="openai",
    mode="honeypot"  # or "tarpit", "redirect"
)

# Analyze any LLM input
result = shield.analyze(user_input)

if result.is_attack:
    # Return deceptive response
    print(result.decoy_response)
else:
    # Safe -- forward to LLM
    response = llm.chat(user_input)
0.98
ML F1 Score
2ms
ML Inference
85-90%
Detection Rate
12
LLM Providers
9
SDK Integrations
280+
Tests

Don't Just Block. Deceive.

Traditional AI firewalls give attackers instant feedback. Oubliette makes them think they succeeded.

The Old Way

  • ✗ Block and return error message
  • ✗ Attacker knows they were detected
  • ✗ Iterates with new attack variants
  • ✗ No intelligence gathered
  • ✗ Single-request analysis only

The Oubliette Way

  • ✓ Return convincing fake data
  • ✓ Attacker wastes hours on decoys
  • ✓ Honey tokens trigger alerts on use
  • ✓ STIX 2.1 threat intelligence generated
  • ✓ Multi-turn session tracking

5-Stage Detection Pipeline

Block obvious attacks in microseconds. Reserve expensive LLM calls for the 5-15% that need them.

1

Input Sanitizer

<1ms

Strips 9 types of encoding attacks, Unicode obfuscation, and invisible characters before any analysis begins.

2

Pre-Filter

~10ms

11 pattern-matching rules block obvious prompt injections, jailbreaks, and DAN attacks instantly. 1,550x faster than LLM-only.

3

ML Classifier

~2ms

LogisticRegression + TF-IDF with 733 features. F1=0.98, AUC=0.99. Catches sophisticated attacks the pre-filter misses.

4

LLM Judge

12 providers

Only 5-15% of inputs reach the LLM judge. Supports OpenAI, Anthropic, Azure, Bedrock, Vertex, Ollama, and more.

5

Session Tracker

multi-turn

Accumulates attack signals across conversation turns. Escalates sessions when thresholds are exceeded.

"Most attacks are obvious — a pattern match catches it in 10 milliseconds. Only the truly ambiguous inputs need the full LLM judge."

Everything You Need to Secure Your LLMs

From input sanitization to threat intelligence, one package handles it all.

4-Tier Detection

Sanitizer, pre-filter, ML classifier, and LLM judge work together in a tiered ensemble pipeline.

2ms ML Inference

LogisticRegression + TF-IDF classifier with 733 features. F1=0.98, AUC=0.99, under 2ms per request.

Cyber Deception

Three deception modes: honeypot (fake data), tarpit (stalling), and redirect (topic steering).

12 LLM Providers

OpenAI, Anthropic, Azure, Bedrock, Vertex, Gemini, Ollama, LiteLLM, Cohere, Mistral, Groq, vLLM.

Multi-Turn Tracking

Accumulates attack signals across conversation turns. Escalates sessions when thresholds are exceeded.

9 SDK Integrations

LangChain, FastAPI, LiteLLM, LangGraph, CrewAI, Haystack, Semantic Kernel, DSPy, LlamaIndex.

CEF/SIEM Logging

ArcSight CEF Rev 25 compliant output to file, syslog, or stdout. Full audit trail for compliance.

Air-Gap Deployment

Zero cloud dependencies with Ollama backend. Runs in SCIF environments with no external connectivity.

Tuned for Precision

Low false positive rate in testing. Benign queries pass through unimpeded while attacks are caught or deceived.

9 Framework Integrations

Drop-in security for every major LLM framework. One pip install away.

LangChain

Callback handler for chains & agents

FastAPI

ASGI middleware for API endpoints

LiteLLM

Universal proxy callback

LangGraph

Shield node for stateful graphs

CrewAI

Task callback + shield tool

Haystack

Pipeline guard component

Semantic Kernel

Prompt filter plugin

DSPy

Assertion module wrapper

LlamaIndex

Callback handler for queries

pip install oubliette-shield[langchain,fastapi,litellm]

Compliance-Ready from Day One

Mapped to every major AI security framework. Audit-ready documentation included.

OWASP LLM Top 10
10/10 categories
OWASP Agentic AI
15/15 categories
MITRE ATLAS
13 techniques
NIST AI RMF 1.0
4 functions
NIST SP 800-53
9 controls
CMMC 2.0
5 domains
CWE
13 identifiers
CVSS v3.1
Severity mapping
NIST CSF 2.0
12 subcategories
Veteran-Owned Small Business

Built for Government & Enterprise

Oubliette Security is a veteran-owned small business pursuing SDVOSB certification. We specialize in AI security for defense, intelligence, and regulated industries — with air-gapped deployment options and full compliance mapping.

NAICS: 541512, 541519, 518210 · SAM Registration in Progress

Defense Solutions Contact Us

Ready to Secure Your AI?

Get started in 3 lines of Python. Free and open source under Apache 2.0.

Get Started Request Demo