Benchling secures multi-tenant AI agents with Amazon Bedrock AgentCore

The solution isolates untrusted code in a dedicated AWS account separate from Benchling’s production environment. Network isolation is enforced with a VPC that has no internet or NAT gateway, a dedicated security group, Route 53 Resolver DNS Firewall (deny‑list, allow‑list, catch‑all), VPC endpoint policies for S3 access, and NACLs that limit traffic to port 443. Per‑job credentials are injected via AWS STS, ensuring each session can only reach the specific tenant’s data bucket. Continuous integration tests simulate DNS tunneling, unauthorized endpoint connections, and out‑of‑scope S3 requests to validate the defenses.

Benchling’s security team chose this managed approach over building a custom sandbox, citing reduced engineering effort and the ability to focus on product security. Jeremy Stashewsky, Application Security Engineer at Benchling, said the pattern lets them “buy instead of build” a secure, scalable AI execution environment.