Benchling deployed Amazon Bedrock AgentCore to secure multi-tenant AI agents, enabling safe execution of scientific code across thousands of life sciences tenants. The solution processes more than 600 code execution sessions per day with zero security incidents.
The architecture uses Amazon Virtual Private Cloud (VPC) mode, combining account-level isolation, Route 5, Resolver DNS Firewall, and VPC endpoint policies to prevent data exfiltration. It enforces strict per-job data access controls without requiring one IAM role per tenant.
"Code Interpreter runs inside a dedicated Security Group restricted to port 443, with no outbound path to the public internet," said Benchling’s security team. This setup blocks unauthorized network connections and ensures data remains isolated within the VPC.
The solution separates untrusted code execution into a dedicated AWS account, isolating it from the main production account. This prevents exposure of customer data and access roles in case of a security breach.
Benchling’s approach includes a continuous validation suite that simulates exfiltration attempts to test the security controls. The DNS Firewall uses a three-priority resolver policy to block known malicious domains and restrict DNS resolution to only explicitly allowed endpoints.
Benchling did not specify the exact number of tenants involved, and the solution’s scalability remains a key consideration for future expansion. The company plans to continue refining the architecture to meet evolving security demands.
Source: awsml
