Abnormal AI Implements Amazon Bedrock AgentCore Code Interpreter for Scalable Email Threat Detection

Abnormal AI, a behavioral security service protecting over a quarter of the Fortune 500, has deployed Amazon Bedrock AgentCore Code Interpreter to support its real-time email threat detection system. This capability provides a fully managed, serverless runtime environment where AI agents can execute code dynamically, process data, and generate insights at an unprecedented scale.

The Code Interpreter operates within ephemeral MicroVM sessions, which can be configured for durations from 15 minutes up to 8 hours, ensuring both flexibility and security. These sessions run in isolated sandboxes with strict separation at the host OS level, preventing data leaks and unauthorized access. The system supports flexible networking options, allowing agents to connect via virtual private clouds or the public internet, and can handle files up to 100 MB directly or connect to Amazon S3 for larger datasets.

Abnormal AI’s architecture employs a three-tiered email detection pipeline, with the most complex cases handled by inline agents utilizing Code Interpreter. These agents analyze threat intelligence data, generate scripts dynamically, and evaluate behavioral patterns to identify malicious activity. The system processes billions of messages daily, with the inline agents focusing on tens of thousands of high-confidence threats.

In addition to real-time detection, Abnormal AI uses a batch analyst agent that ingests misclassifications and patterns, running long-duration Code Interpreter sessions to improve heuristics and models. This approach enables continuous learning and refinement of detection capabilities.

Security considerations led Abnormal AI to adopt a no-egress sandbox configuration, ensuring data reproducibility and preventing exfiltration even if the agent becomes malicious. Additional security layers include controlled data ingestion, subprocessor compliance, and network isolation.

Practical lessons from this deployment emphasize giving agents autonomy with high-level principles, providing compute scratch pads for data analysis, employing programmatic verifiers for output quality, and using file systems as checkpoints for long-running tasks. These practices enhance the robustness and security of AI agent systems.

Abnormal AI’s implementation demonstrates that Code Interpreter is more than a coding tool; it is a fundamental infrastructure component that enables agents to reason computationally. By combining secure, managed sandbox environments with lightweight agent harnesses, organizations can scale AI security solutions effectively while maintaining strict security standards.