Delphi Security Launches xAIDR: First Runtime Benchmark for Agent-to-Agent Attack Detection

Introduction

Delphi Security is a company focused on protecting artificial intelligence systems. Recently, they introduced a new tool called xAIDR, which stands for Extended AI Detection and Response. This tool is designed to find and stop attacks that happen when one AI agent talks to another. It is the first published benchmark specifically built for this type of security challenge.

Benefits

The main advantage of xAIDR is its ability to catch threats that traditional security tools miss. Most security software only watches the end points where users interact with AI. xAIDR works at the runtime layer, meaning it checks messages while they are being sent between agents. This catches problems before they can cause harm.

The tool has shown strong results in testing. It achieved a detection accuracy of 94.5 percent and a precision of 98.4 percent. It works with agents from many different companies, including OpenAI, Anthropic, Gemini, Groq, and Azure. This means it does not depend on a single vendor, making it a flexible solution for complex systems.

xAIDR covers many dangerous attack types. These include prompt injection, which is currently the top security risk for large language models. It also handles issues like agent identity spoofing, goal hijacking, trust escalation, and data theft. By addressing these specific threats, it helps keep multi-agent systems safe.

Use Cases

This tool is useful for any organization that uses multiple AI agents working together. As AI systems become more autonomous, they often need to communicate with each other to complete tasks. This communication creates new security risks because agents from different vendors might not be able to protect each other.

xAIDR is ideal for environments where agents from different companies interact. It can intercept threats within these messages before they are executed. This is critical for businesses building complex workflows where one agent might need to pass information to another. It ensures that sensitive data does not leak and that agents do not follow malicious instructions sent by other agents.

Pricing

Pricing details for Delphi Security and xAIDR are not available in the provided information.

Vibes

Public reception highlights the importance of this new tool. Experts note that the emergence of xAIDR is a significant step forward in securing multi-agent AI systems. The vendor-agnostic approach is seen as a potential new standard for runtime AI security. However, some experts recommend independent third-party validation to confirm the reported accuracy figures. They also suggest that real-world results will depend on how well the tool works in production environments beyond the controlled tests used during development.

Additional Information

Delphi Security launched xAIDR on April 23, 2026. The benchmark tested 500 different adversarial scenarios across 12 attack categories. This testing proved that vendor-agnostic runtime detection is both necessary and technically achievable for inter-agent communications.