The Platform
Six packages shipping on PyPI today, one unified install, and a Phase I agentic prototype available under briefing.
Oubliette Shield
Five-stage runtime detection pipeline that blocks prompt injection and jailbreak attacks, then deceives attackers with honeypots and honey tokens. Pre-filter decisions in under a millisecond (0.006 ms median, 2026-04-22 throughput run). 12 LLM-judge providers, 9 framework integrations.
$ pip install oubliette-shield Oubliette Dungeon
Multi-agent red team for AI systems with 72 YAML-defined attack scenarios (57-scenario core library + 15 Crescendo multi-turn), 10 categories. Multi-provider comparison, React review dashboard, MCP server, polyglot tool integration.
$ pip install oubliette-dungeon Oubliette Trap
AI-agent deception platform — 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.
$ pip install oubliette-trap Oubliette Warden
Safety-gated, human-on-the-loop multi-agent framework for authorized cyber operations. The planner turns an objective into an ATT&CK-aligned task graph; every tool invocation then passes the Shield safety pipeline and a plan-consistency check before an operator approves it. The gate evaluates and executes nothing on its own.
$ pip install oubliette-warden Oubliette Barbican
Defensive detector for coordinated synthetic influence content in short-form social posts. Coordination-first: per-post authenticity signals feed a cross-artifact correlator, because scoring one post at a time overstates defense. Zero runtime dependencies, and the bundled demo corpus runs on a disconnected host.
$ pip install oubliette-barbican Oubliette Suite
Umbrella package with a unified CLI and a threat-model scoping engine that turns a few questions about your system into a run plan. Install by outcome — [assess], [defend], [deceive], [detect] — or [full] for everything. The fastest way to evaluate the platform end to end.
$ pip install "oubliette-suite[full]" Agentic Cyber Defense
Four-agent framework (Project Management, Code Generation, Cyber Analysis, Vulnerability Research) with a human-on-the-loop operator interface. CALDERA-emulated; citation-bound vulnerability research. Shield and Dungeon map findings to MITRE ATLAS (v2026.06); Warden plans are ATT&CK-aligned.
How the pieces fit
Each capability is independently useful. They reinforce each other when deployed together.
Shield ↔ Dungeon
Dungeon attacks your AI under controlled scenarios. Shield defends. Measure the detection rate. Close the gaps. Repeat.
Agentic ↔ Shield
Shield and Dungeon map findings to MITRE ATLAS (v2026.06); Warden plans are ATT&CK-aligned. No black-box autonomy.
Agentic ↔ AI RMF
The agentic framework's evidence-integrity pattern binds every vulnerability claim to a retrieved corpus record — citations grounded, not synthesized. Aligned with NIST AI RMF GOVERN / MAP / MEASURE / MANAGE controls.
Trap ↔ Shield ↔ Dungeon
Trap's honeypot MCP servers fingerprint the agents that probe you while Shield's deception gateway feeds them decoys. What attackers actually do inside becomes new Dungeon scenarios and new Shield rules — a loop a detection-only stack cannot close.