Project Black
Monolith.
A unified security middleware system designed to intercept, filter, and ground autonomous agent executions.
Unified Shield
Black Monolith is a unified agentic AI security middleware system built to address a growing problem. As AI agents get more autonomous and tool-using (via MCP and similar protocols), there's no unified layer watching what they do, what data they touch, or whether their reasoning can be trusted in the moment. Black Monolith is an attempt to build that layer.
The architecture operates on a zero-trust model, forcing all agent tool invocations through dynamic sanitization, grounding, and logical oversight modules before code execution.
The Problem
Most existing security tooling was built for static workflows and human-in-the-loop triggers. It was never designed for agents that reason dynamically, call tool suites autonomousy, and access sensitive database structures in fractions of a second.
Traditional defenses are largely reactive. They log events, generate audit traces after the fact, and alert developers once an exploit has already executed. This approach fails when an agent goes off-rails or suffers prompt injection in the middle of a live production database query.
To safely deploy autonomous systems, we need active interception through a middleware layer capable of inspecting, modifying, and blocking actions in real time, before execution commands ever reach their targets.
Modular Defenses
MCP-Shield
Rust ProxyA high-performance proxy layer sitting between agents and their MCP tool calls. Intercepts every tool invocation before execution, evaluates it against configurable policy rules, and can block, modify, or flag calls in real time. Built in Rust for low-latency interception that doesn't bottleneck agent performance.
VectorAnchor
FastAPI + RAG MiddlewareA grounding layer that sits between an agent's reasoning and its outputs, cross-referencing responses against a trusted knowledge base in real time. Designed to catch hallucinated tool calls, fabricated data references, and ungrounded claims before they reach the user or downstream systems.
TraceAudit
Streaming MonitorA real-time observability layer that streams and analyzes an agent's reasoning trace as it happens, rather than after the fact. Flags anomalous reasoning patterns, unexpected tool-use sequences, or signs of prompt injection / goal drift, surfacing them to a live dashboard instead of a static log.
System Architecture
The core components of Black Monolith connect to form a secure execution boundary. MCP-Shield acts at the tool-call layer, intercepting outbound requests, while VectorAnchor grounds the output layer, cross-referencing model reasoning against a verified context buffer.
Above both systems, TraceAudit provides continuous, stream-based observation of the internal reasoning trajectory. All three components send diagnostic and telemetry events directly to a unified dashboard, enabling operators to inspect and intervene in real time.
Dashboard Integration
All telemetry data generated by MCP-Shield, VectorAnchor, and TraceAudit feed into a unified Next.js 15 dashboard. This provides full streaming transparency, latency metrics, and instant intercept capability.
Quick Reference & Answers
What is Project Black Monolith?
Project Black Monolith is a unified security middleware suite developed by Sleepers Research. It acts as an active interception layer between AI agents and external environments, monitoring, filtering, and grounding tool executions in real-time.
How does MCP-Shield secure agentic tool calls?
MCP-Shield operates as a high-performance proxy layer written in Rust. It intercepts Model Context Protocol (MCP) tool invocations before execution, evaluating each call against customizable runtime policy rules to block, modify, or alert on security violations with sub-millisecond overhead.
What is the role of VectorAnchor in grounding LLM responses?
VectorAnchor functions as a RAG (Retrieval-Augmented Generation) middleware layer. It cross-references agent outputs and tool calls against a verified database context in real-time, catching hallucinated data, invalid parameters, or unauthorized actions before they reach execution layers.
“As agents get more capable, the layer watching them has to get smarter too.”