Delphi Security Launches xAIDR: First Runtime Benchmark for Agent-to-Agent Attack Detection
Key Takeaways
- ▸xAIDR achieves 94.5% detection accuracy with 98.4% precision across 500 adversarial A2A attack scenarios spanning 12 categories
- ▸First published benchmark demonstrating that vendor-agnostic runtime detection is both necessary and technically achievable for inter-agent communications
- ▸Covers critical attack vectors including prompt injection, identity spoofing, goal hijacking, and MCP poisoning against major LLM platforms including OpenAI, Anthropic, and Gemini
Summary
Delphi Security has introduced xAIDR (Extended AI Detection & Response), marking the first published runtime benchmark for detecting and responding to attacks occurring within agent-to-agent (A2A) communications. The benchmark evaluates 500 adversarial A2A scenarios across 12 attack categories, achieving 94.5% detection accuracy and 98.4% precision. The testing was conducted across agents powered by major AI vendors including OpenAI, Anthropic, Gemini, Groq, and Azure, demonstrating vendor-agnostic detection capabilities.
Unlike traditional security tools that monitor AI agent endpoints, xAIDR operates at the runtime layer to intercept threats within inter-agent messages before execution. The benchmark covers critical attack vectors including prompt injection via A2A, agent identity spoofing, goal hijacking, trust escalation, MCP tool poisoning, data exfiltration, unauthorized delegation, and memory poisoning. With prompt injection ranked #1 in OWASP's LLM Top 10 for 2025-2026, Delphi Security argues that existing security solutions remain blind to cross-vendor A2A attacks that vendors on either end cannot independently detect.
- Addresses a critical security gap where traditional endpoint-focused tools cannot detect attacks occurring between agents from different AI vendors
Editorial Opinion
The emergence of xAIDR represents a significant step forward in securing the increasingly complex landscape of multi-agent AI systems. As AI agents become more autonomous and interact with each other across vendor boundaries, the security blind spots in existing tools become more critical—and xAIDR's vendor-agnostic approach could establish a new standard for runtime AI security. However, the claim of 94.5% detection accuracy should be validated by independent third parties, and the real-world applicability of this benchmark will depend on how well it generalizes to production environments beyond the controlled scenarios tested.
