Agent Armor

Agent Armor is an open-source security framework designed to detect and defend against a wide range of threats targeting AI agents. The tool addresses risks such as content injection, jailbreak patterns, RAG poisoning, data exfiltration, unauthorized sub-agent spawning, and semantic manipulation, among others. It is built to scan inputs, RAG (Retrieval-Augmented Generation) chunks, and outputs before they reach large language models, aiming to prevent adversarial attacks and manipulation in AI agent systems.

The framework offers middleware that can be integrated into AI workflows, enabling developers to scan and sanitize user inputs as well as RAG chunks prior to context assembly. Agent Armor’s detection capabilities include identifying hidden HTML, metadata injection, dynamic cloaking, syntactic masking, oversight evasion, and multi-turn session attacks where malicious instructions are distributed across conversation turns. The tool employs both regex-based detection patterns and an optional machine learning classifier based on the DeBERTa-v3 multi-label model via ONNX, providing asynchronous detection and closing gaps that traditional pattern matching may miss. Evaluation results on curated adversarial and benign samples indicate high detection rates with minimal false positives across different strictness levels.

Agent Armor is implemented as a JavaScript package that can be installed via npm, with usage demonstrated through code examples for scanning and sanitizing text and RAG chunks. The open-source project is licensed under the MIT License and is developed by Stylus Nexus Holdings, LLC. Its threat taxonomy and detection strategies draw on the DeepMind Agent Trap Taxonomy, situating it within a class of tools focused on AI agent observability and security.

evidence_sufficient": true}