Abnormal AI: Amazon Bedrock AgentCore for agentic email security at scale

This article describes how Abnormal AI uses Amazon Bedrock AgentCore Code Interpreter to power agentic email security systems processing billions of messages daily.

• Code Interpreter provides managed, serverless runtime for agents to execute code in secure, isolated sandboxes with configurable session durations up to 8 hours
• Abnormal AI's three-tiered detection pipeline uses heuristics for billions of messages, ML models for millions, and inline agents with Code Interpreter for tens of thousands of hardest cases
• Analyst agent operates in batch mode to identify patterns, autonomously write heuristics, and improve models based on misclassifications from the real-time pipeline
• Zero-trust sandbox design prevents data exfiltration and ensures reproducibility by isolating agents from external network access
• Best practices include giving agents lightweight harnesses, providing compute scratch pads for all agents, using programmatic verifiers as guardrails, and checkpointing state for long-running tasks

Code Interpreter serves as fundamental infrastructure for production agents to perform computational reasoning beyond semantic analysis, enabling secure threat detection at scale.