EXTENDED AI DETECTION AND RESPONSE
The security layer for multi-agent AI systems
Every inter-agent message scanned inline. Fleet behavioral correlation. Shadow agent discovery. Trust-based autonomous governance. Under 20ms.
WHAT IS xAIDR
xAIDR, Extended AI Detection and Response, is a runtime security architecture for multi-agent AI systems that provides inline content scanning of all inter-agent messages, fleet-level behavioral correlation across agents, autonomous discovery of unregistered agents, and trust-based governance with automated response.
An xAIDR system must operate at runtime, not at deployment time. It must scan content, not just metadata. It must reason across agents, not just within them.
FOUR CAPABILITIES
What xAIDR does
Inline Content Scanning
Every inter-agent message scanned before it reaches the LLM. Under 20ms. No network hop. No proxy round-trip. Functions even when Sentinel Brain is unreachable.
Fleet Behavioral Correlation
When an attacker probes three agents in 90 seconds, Sentinel connects the dots. Three blocked events become one coordinated attack chain alert, not three separate noise events.
Shadow Agent Discovery
When a registered agent receives a message from an agent nobody deployed, Sentinel flags it instantly and automatically. No manual inventory required.
Trust-Based Governance
Every agent carries a live trust score that degrades on anomalous behavior. When it drops below threshold, Sentinel quarantines the agent automatically. No ticket. No delay. No human required.
THE BENCHMARK
94.5% A2A detection accuracy
First published benchmark for agent-to-agent runtime detection. 5 iterative evaluation runs on a 500-prompt A2A test suite.
Run 1
Baseline detection rules
81.6%
Run 2
A2A pattern library added
86.3%
Run 3
Intent decomposition added
90.7%
Run 4
Tool call anomaly detection
93.1%
Run 5
Chat-mode agentic rules
94.5%
98.4%
PRECISION
92.9%
RECALL
500
TEST PROMPTS
ARCHITECTURE
Two components. Complete coverage.
The Sensor
A lightweight library embedded in each agent. Intercepts every message before it reaches the LLM. Runs L1 and L2 detection locally. Adds under 20ms. Works even when the central brain is unreachable.
Sentinel Brain
Centralized intelligence engine with real-time awareness of your entire agent fleet. Receives telemetry from every sensor. Correlates behavior across agents. Maintains trust scores. Triggers autonomous quarantine.
MICROSOFT AGT INTEGRATION
Open governance standard. Delphi enforcement.
Delphi's sensor wraps Microsoft's Agent Governance Toolkit components including PolicyEngine, AgentIdentity using Ed25519 and DID, and circuit breakers as internal dependencies. Install xaidr and get both Delphi content inspection and AGT-compliant governance in one package.