Benchling Secures Multi-Tenant AI Agents with Amazon Bedrock AgentCore
Summary:Benchling has implemented a robust security architecture for multi-tenant AI agents using Amazon Bedrock AgentCore Code Interpreter, Amazon Route 53 Resolver DNS Firewall, and VPC endpoint policies. This setup ensures complete network isolation for AI-generated code, preventing data exfiltration and safeguarding tenant data. Since its deployment in April 2026, Benchling has processed over 600 code execution sessions daily across more than 250 tenants per week without any security incidents. This
Background and Challenges
Benchling, a technology company providing AI platforms for the life sciences sector, needs to generate and execute scientific code on behalf of thousands of tenants. Traditional sandboxing techniques proved inadequate in the face of complex network threats, especially when AI agents handle sensitive data, where the risk of data exfiltration is particularly high. Benchling's security team recognized that standard network controls, such as restricting HTTP requests and egress ports, were insufficient to meet their security requirements, particularly the lack of control over DNS resolution, which could lead to data breaches.
Solution Architecture
Benchling adopted Amazon Bedrock AgentCore Code Interpreter and deployed it in Amazon VPC mode, integrating the following security measures to build a multi-layered security architecture:
- Account-level Isolation: AI-generated code runs in a separate AWS account, completely isolated from production systems, mitigating potential security risks.
- Amazon Route 53 Resolver DNS Firewall: Utilizes a three-tier priority policy to block malicious DNS resolutions:
- P10 (High Priority Block): Blocks known malicious domains.
- P100 (Allow List): Only allows DNS resolutions for explicitly listed domains, primarily S3 bucket endpoints.
- P200 (Block All): Blocks all domains not listed in the allow list, preventing DNS tunneling attacks.
- VPC Endpoint Policies: Controls access to AWS services through Gateway and Interface VPC endpoints, allowing only specific S3 bucket access requests.
- Network Access Control Lists (NACLs): Restricts traffic to only port 443 and ephemeral return ports, further reducing the attack surface.
- Per-Job Credential Scoping: Injects temporary credentials into each job through AWS STS, ensuring that code can only access data for a specific tenant.
Key Technologies and Advantages
- DNS Firewall Configuration: The three-tier policy of the Route 53 Resolver DNS Firewall effectively blocks DNS tunneling attacks and provides detailed logs for security audits.
- VPC Endpoint Policies: Fine-grained endpoint policies allow Benchling to precisely control access to AWS services, preventing unauthorized data access.
- Per-Job Credential Scoping: Avoids pre-provisioning IAM roles for each tenant by dynamically injecting temporary credentials, reducing the complexity of IAM role management.
Continuous Validation and Testing
Benchling's team uses a continuous integration test suite to simulate various data exfiltration scenarios, including DNS tunneling attacks, direct connections to unauthorized endpoints, and attempts to access unauthorized S3 buckets. These tests ensure the effectiveness of the security configuration and help detect potential vulnerabilities as the environment evolves.
Business Impact
Since its deployment in April 2026, Benchling has successfully processed over 600 daily code execution sessions across more than 250 tenants per week without any security incidents. This solution not only enhances the security of AI agents in multi-tenant environments but also frees up resources for Benchling's core product development.
Conclusion
Benchling's case demonstrates that by combining Amazon Bedrock AgentCore, VPC mode, and fine-grained network control strategies, AI agents can operate securely in multi-tenant environments. This solution provides a secure, scalable architecture blueprint for AI-driven life science applications.
— END —Source: AWS Machine Learning Blog (2026-09-21)
Tags: #Benchling #AI Security #Multi-Tenant Architecture #Amazon Bedrock #VPC
Editorial & Fact-Checking Note: This article is compiled from primary research, official release documentation, and source papers by the ZICQ Newsroom pipeline with automated entity verification and human editorial review. If you notice any technical inaccuracy, please submit a correction via our corrections policy or email our editorial desk directly.
Community Comments