Pratyahara: A Neural Tissue Defense Model for Detecting Compromised Agents in Multi-Agent Networks

Published April 17, 2026 | Version v1

Preprint Open

Multi-agent networks built on the Agent2Agent (A2A) protocol [5] and Model

Context Protocol (MCP) [6] have a gap that perimeter defense cannot close:

agents whose behavior drifts over time due to adversarial compromise,

reinforcement learning misalignment, self-healing side effects, or model

update artifacts. Research shows that a majority of large language models

can be manipulated through inter-agent communication, with attack success

rates exceeding 80% before mitigation [1]. Session smuggling [2], tool

poisoning [3], and emergent collusion all operate inside the network

boundary where input filtering is structurally blind.

We propose the NERVE specification, five primitives for detecting and

responding to behavioral drift in multi-agent networks. The design draws

on the brain's neural tissue defense mechanisms, which are intent-agnostic:

they detect deviation from baseline regardless of whether the cause is

malicious, accidental, or emergent. NERVE extends the A2A and MCP

protocols without requiring changes to either core specification and is

analyzed against eight documented attack vectors.

Files

Name	Size
pratyahara-nerve-protocol.pdf md5:e6aebdb11fd6bf6414f9e5303d4d5c4a	443.9 kB	Preview Download