Cyber, AI, and Agentic Security
for the Defense Enterprise
Detect, validate, and deceive adversarial threats across AI and agentic systems. Five shipping products, built by a Service-Disabled Veteran-Owned Small Business.
One Suite. Five Products.
Each product is independently useful — together they form the assess / defend / deceive / detect stack for AI and agentic systems in contested environments, with an agentic prototype available under briefing.
Runtime AI Firewall
Five-stage detection pipeline — sanitizer, pre-filter, ML classifier, RAG guard, and LLM judge — gates every input to your LLM. Pre-filter decisions in under a millisecond (0.006 ms median, 2026-04-22 throughput run); inputs that go to the LLM judge take longer. Cyber-deception mode turns blocks into honeypots so attackers waste hours on decoys while you collect threat intelligence.
Adversarial Testing Engine
Multi-agent red team for AI systems with 72 attack scenarios (57-scenario core library + 15 Crescendo multi-turn), 10 categories. React review dashboard, MCP server, polyglot tool integration (Garak, PyRIT, AI Exchange). Findings mapped to MITRE ATLAS (v2026.06), OWASP LLM Top 10, and NIST AI RMF.
Deception Decoys
Honeypot MCP servers that attract, contain, and fingerprint autonomous AI agents inside interconnected fake environments. Behavioral classification (LLM, script, human, compromised) with STIX 2.1 / CEF intelligence export — the deception half a detection-only stack does not have.
Agent Containment
Safety-gated, human-on-the-loop framework for authorized cyber operations. ATT&CK-aligned plans, every tool call gated by the Shield pipeline and a plan-consistency verifier, and an operator approval step before anything runs.
Influence Detection
Defensive detector for coordinated synthetic influence content aimed at you. Coordination-first: per-post authenticity signals feed a cross-artifact correlator. Zero runtime dependencies; the demo corpus runs air-gapped.
Agentic Cyber Defense
A four-agent framework (Project Management, Code Generation, Cyber Analysis, Vulnerability Research) with a human-on-the-loop operator interface. Every autonomous action gated by the same safety pipeline that powers Shield. CALDERA-emulated; ATT&CK-aligned; citation-bound vulnerability research.
Shield, Dungeon, Trap, Warden and Barbican are public PyPI packages you can install and test: pip install "oubliette-suite[full]". The multi-agent framework is a prototype and is not in the package. Shield and Dungeon map findings to MITRE ATLAS (v2026.06); Warden plans are ATT&CK-aligned.
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 under a millisecond. Reserve expensive LLM calls for the inputs that need them.
Input Sanitizer
<1msStrips 9 types of encoding attacks, Unicode obfuscation, and invisible characters before any analysis begins.
Pre-Filter
<1ms11 pre-filter rules block obvious prompt injections, jailbreaks, and DAN attacks. Sub-millisecond: 0.11 ms mean in the 2026-03-12 comparison run.
ML Classifier
733 featuresServer/demo model: LogisticRegression + TF-IDF with 733 features. Weighted F1 0.98 / ROC-AUC 0.99 on a held-out split of template-generated training data; not yet evaluated on an external dataset. Catches sophisticated attacks the pre-filter misses.
LLM Judge
12 LLM-judge providersOnly inputs the earlier stages cannot settle reach the LLM judge. Supports OpenAI, Anthropic, Azure, Bedrock, Vertex, Ollama, and more.
Session Tracker
multi-turnAccumulates attack signals across conversation turns. Escalates sessions when thresholds are exceeded.
"Most attacks are obvious — a pattern match catches it in under a millisecond. Only the truly ambiguous inputs need the full LLM judge."
Compliance-Ready from Day One
Mapped to every major AI security framework. Audit-ready documentation included.
Built for Government & Enterprise
Oubliette Security is an SBA-certified Service-Disabled Veteran-Owned Small Business (SDVOSB). We specialize in AI security for defense, intelligence, and regulated industries — with air-gapped deployment options and full compliance mapping. Eligible for SDVOSB sole-source awards up to $5M and set-aside competitions.
NAICS: 541512, 541519, 518210 · SAM.gov Registered · CAGE: 19AK6
Ready to Engage?
Federal program offices, AI builders, and incident responders — three different ways to start with Oubliette Security.