How Benchling secured multi-tenant AI agents with Amazon Bedrock AgentCore

This article describes how Benchling built a defense-in-depth security architecture to safely run AI agent-generated scientific code across thousands of life sciences tenants using Amazon Bedrock AgentCore.

• Deployed AgentCore Code Interpreter in VPC mode with account-level isolation separating untrusted code from production systems
• Implemented Route 53 Resolver DNS Firewall with three-priority policy: block malicious domains, allow only explicit endpoints, deny all others
• Used VPC endpoint policies and NACLs to restrict network access to specific S3 buckets and port 443 only
• Injected per-job scoped credentials via AWS STS to limit data access to individual tenants without creating role sprawl
• Established continuous validation test suite simulating exfiltration attempts to ensure security posture remains intact
• Achieved 600+ daily code execution sessions across 250+ tenants per week with zero security incidents since April 2026

The multi-layered approach combining account isolation, DNS firewall controls, VPC endpoint policies, and continuous validation provides a blueprint for secure multi-tenant code execution without building custom sandboxing infrastructure.