OWASP Foundation web repository

Open-source governed memory for AI agents: prompt-injection-resistant writes, purpose-bound retrieval, provenance, and tamper-evident audit.

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

Antivirus for AI agent memory - scanner + MCP firewall for memory poisoning (OWASP ASI06)

AI Agent Security — Attack payloads, defense references, and research. 52 tests, ~10K lines. A learning-oriented shooting range, not a product.

AI Security SDK for Agents, LLMs & SLMs

Stop memory poisoning attacks on your AI agents

A drop-in write gate for agent memory. Wrap your Mem0 client; block memory poisoning at the write path.

Agentic AI Request Forgery (AARF) – New vulnerability class exploiting planner ➝ memory ➝ plugin chaining in MCP Server, MAS, LangChain, and A2A agents. Red Team playbooks, threat models, OWASP Top 10 proposal.

Integrated memory-poisoning defense for persistent LLM agents (provenance + lineage + trajectory detection). AGPL-3.0.

Ratine — Agent memory poisoning detector. Scans AI agent persistent memory for injected instructions, hidden payloads, credential leakage, and belief drift. OWASP ASI06. Zero dependencies.

Durable, private, time-aware memory engine for long-running AI agents

Protect AI agent memory from poisoning attacks with a zero-dependency shield that fits Mem0, LangChain, or custom memory systems.

ZKP-RA: Zero-Knowledge Proof Reasoning Anchors for mitigating memory poisoning in autonomous agentic DeFi — Groth16 zk-SNARKs, Circom circuits, on-chain Solidity verifier, Python agent guardian. 182ms proofs, 9.7x gas reduction vs zkML.

This repository documents AI Recommendation Poisoning — a real-world attack technique discovered by Microsoft's Defender Security Research Team (February 2026) where adversaries silently inject persistent instructions into AI assistant memory through carefully crafted URLs.

Reproducible benchmark for memory poisoning in persistent AI agents: can untrusted content silently become trusted knowledge or operating policy? 7 attack categories, per-trial evidence.

Does poisoned long-term memory make frontier LLMs misbehave? A rigorous AI-safety eval across 7 risks × 4 models (Claude Opus 4.8 & Sonnet 4.6, GPT-5.5, Gemini 3.5) with prompt-level and dose-response analysis.

AeroMind: Poisoning the Control Plane of LLM-Driven UAV Agents (RAID 2026)

An open trust layer for AI-agent memory and identity: tamper-evident memory, cryptographic agent identity (SoulKeys), capability-scoped governance, and OWASP-ASI06 memory-poisoning detection. Apache-2.0.

Memory as a Control Plane: Poisoning Attacks on LLM Multi-Agent UAV Systems - IEEE Conference Paper