Benchling says it is running AI-generated scientific code for thousands of life-sciences tenants inside a deliberately sealed AWS environment, using DNS controls, scoped credentials and continuous exfiltration tests. The important point is not that someone has built another sandbox. It is that the company treats the network boundary itself as a security product, rather than trusting the code to behave.
AWS AI Watch analysis
What happened
In an AWS machine-learning blog post published on 21 September, AWS describes Benchling’s use of Amazon Bedrock AgentCore Code Interpreter in a separate AWS account for untrusted execution. Benchling says the system processes more than 600 code-execution sessions a day across more than 250 tenants each week, with no security incidents reported by the company.
The design keeps the untrusted execution account apart from Benchling’s production account and customer data. There is no internet gateway or NAT gateway. Code runs inside a VPC with restricted security groups, VPC endpoints and Route 53 Resolver DNS Firewall rules. DNS requests are blocked unless they match an explicitly allowed list, while endpoint policies restrict which S3 buckets a job can reach.
Benchling injects short-lived, per-job credentials through AWS Security Token Service rather than creating a permanent IAM role for every tenant. Its security team also says it tests DNS tunnelling, direct connections to unauthorised endpoints and attempts to access unapproved S3 buckets. Those checks have been added to continuous integration so infrastructure changes can fail before release.
Key findings
- Separate execution account
AI-generated code is kept away from the main production account and broader customer data store. - DNS deny by default
Route 53 Resolver DNS Firewall blocks listed hostile domains, allows approved endpoints and returns no data for everything else. - Per-job access
Temporary credentials limit each execution session to the data needed for that task. - Continuous attack tests
Benchling repeatedly tests common exfiltration paths instead of treating the initial configuration as finished work.
Why it matters
AI agents that generate and execute code are useful precisely because they can do more than produce text. That is also the problem. A generated script may contain an accidental mistake, follow hostile instructions in its inputs or actively try to move data somewhere it should not go.
Benchling’s approach offers a practical pattern for organisations handling sensitive scientific information: isolate the workload, make outbound access impossible by default, grant only the required data, and test the controls as the infrastructure changes. The useful lesson is less “AWS has solved agent security” than “agent security needs several independent locks”. One clever sandbox is not a strategy.
The evidence remains Benchling’s account of its own architecture and testing. The post does not establish that the design is invulnerable, nor does a reported absence of security incidents prove that every future deployment will behave the same way.
Our read
This is a proper technology story because it shows how an AI system can be made to operate inside a meaningful blast-radius limit. The unglamorous details are the story: no default internet route, tightly scoped storage access and tests designed to catch configuration drift.
Teams deploying agents around private data should borrow the principle, not copy the diagram blindly. The exact controls depend on the workload, cloud account structure and threat model. Security architecture is not a magic spell, sadly, although “deny all by default” is about as close as infrastructure gets to one.
What to watch
- Whether Benchling publishes measurable results from its continuous exfiltration tests.
- How the design handles new services, endpoints and cross-region data access.
- Whether other life-sciences companies adopt separate accounts for untrusted agent execution.
- Independent evidence about the operational cost and failure modes of the pattern.
Discussion spark: Should organisations running AI-generated code make deny-by-default network isolation a baseline requirement, even when it adds friction and cost, or is that too heavy for most agent workloads?
Sources and evidence
- How Benchling secured multi-tenant AI agents with Amazon Bedrock AgentCore (21 September 2026, 16:27 UTC)
not affiliated with or endorsed by Amazon Web Services (AWS)