Project Black
Monolith.
A unified agentic AI security middleware system designed to intercept, filter, and ground autonomous tool executions in real time.
Research Problem
As language models transition from chat interfaces into autonomous agents capable of calling Model Context Protocol (MCP) servers, databases, and APIs, traditional security boundaries disintegrate. Agents operate dynamically: they form variable execution plans, compose multi-step tool calls, and digest uncurated external inputs.
Current security paradigms rely on post-hoc logging or static regex filters. They fail when an agent encounters indirect prompt injection embedded within an external document, leading the agent to generate valid tool calls that execute destructive database mutations or unauthorized data exfiltration.
Why It Matters
Without a dedicated inline security layer, organizations cannot safely grant autonomous agents access to sensitive production APIs. A single ungrounded hallucination or injected instruction can compromise critical systems in fractions of a second.
Research Objective
Develop a zero-trust, low-latency middleware framework that inspects tool arguments, cross-references claims against trusted context, and streams real-time reasoning telemetry without degrading agent response times.
Methodology & Subsystems
MCP-Shield
Rust Proxy LayerA high-performance proxy layer sitting between agents and their Model Context Protocol (MCP) tool calls. Intercepts every tool invocation before execution, evaluates parameters against configurable policy rules, and can block, modify, or flag calls in sub-millisecond latency ranges.
VectorAnchor
FastAPI + RAG MiddlewareA real-time grounding layer that sits between an agent's reasoning trajectory and its external actions, cross-referencing tool arguments and outputs against trusted vector context buffers to prevent ungrounded claims or hallucinated parameter values.
TraceAudit
Streaming Reasoning MonitorA real-time observability subsystem that streams and analyzes an agent's thought process as tokens are generated. Detects anomalous reasoning steps, goal drift, and prompt injection patterns before harmful commands execute.
Key Findings & Outcomes
Initial benchmarks on prototype nodes demonstrate that inline Rust proxying (MCP-Shield) adds less than 4.2ms of overhead per tool call, while VectorAnchor reduces ungrounded tool parameter execution by up to 94% in synthetic prompt injection tests.
Related Research
Reasoning Trace Analysis & Early Failure Prediction
Examines intermediate model thought paths to detect hallucination markers before tool calls reach execution.
Examine Project Undertow Project FathomMechanistic Interpretability of Small Models
Investigates parameter matrix shifts and activation distributions in fine-tuned edge models used inside proxies.
Examine Project FathomReferences & Citations
- Model Context Protocol (MCP) Specification — Model Context Protocol Open Standard, Anthropic (2024).
- Zero-Trust Security Architecture for Autonomous Agents — Sleepers Research Lab Technical Note (2026).
- Process-Supervised Reward Models & Reasoning Trajectories — OpenAI Research & Alignment Literature (2023-2025).
Frequently Asked Questions
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.
Interested in Project Black Monolith?
Reach out to the laboratory to discuss security middleware audits or custom infrastructure integration.
Contact the Laboratory