Abnormal AI is using Amazon Bedrock AgentCore Code Interpreter to give its email-security agents a temporary environment for data analysis, verification and dynamically generated code. The agents support inline threat detection intended to block malicious messages before they reach an inbox.
The security company says its systems process billions of messages and agent-driven operations at that scale. Abnormal protects more than 25% of Fortune 500 companies, according to AWS. Those figures come from the companies involved rather than an independent performance audit.
AgentCore runs each code session in an ephemeral microVM, with host-level separation between sessions. A session lasts 15 minutes by default and can be configured for as long as eight hours. Networking can be restricted to a virtual private cloud or opened to the public internet, and files up to 100 MB can be sent through the API.
A disposable scratch pad reduces the risk of letting model-generated code run directly in a long-lived application environment, but sandboxing does not make the code or its conclusions automatically safe. Builders still need minimal network permissions, controlled inputs and outputs, short retention, and monitoring. The deployment illustrates how production agents increasingly need conventional compute isolation in addition to language-model guardrails.