AgentTrust: Runtime Interception Layer Catches Unsafe AI Agent Tool Calls Before Execution
arXiv·medium signal
AgentTrust proposes a runtime safety evaluation and interception system for AI agent tool use — file operations, shell commands, HTTP requests, and database queries are all intercepted before execution. Unlike post-hoc benchmarks or static guardrails, it understands multi-step context and obfuscation. Arrives alongside Microsoft's Agent Governance Toolkit and AEGIS, signaling runtime agent safety is becoming a crowded, critical space.