GitHub Repository

Memory defense for AI agents — stops MINJA, AgentPoison, and MemoryGraft attacks. Zero dependencies.

4 starsTypeScript

Mguard – First defense against MINJA memory poisoning attacks

Name: Mguard – First defense against MINJA memory poisoning attacks
Availability: InStock
Author: mguardai

by mguardai·Mar 9, 2026·1 point·0 comments

AI Analysis

●●SolidBig BrainWizardryNiche Gem

Ed25519 provenance plus Bayesian trust scoring stops published NeurIPS memory poisoning attacks zero-dep.

Strengths

•Addresses real published academic threat (MINJA 95%+ success rate, NeurIPS 2025) with verifiable defense mechanism before industry built defenses.
•Six-layer defense combining cryptographic signing, trust models, anomaly detection, and pattern matching shows depth beyond snake oil.
•Drop-in wrappers for Mem0 and LangChain lower adoption friction; zero dependencies minimize supply chain risk.

Weaknesses

•No independent security audit or formal verification; claims rely on academic attack papers but defense itself unvalidated by third party.
•Unclear threat model scope: does not address context window attacks, prompt injection variants, or memory exfiltration beyond poisoning patterns.