Securing agents in Coder
Note
This functionality is in early access and still evolving. For now, we recommend testing it in a demo or staging environment, rather than deploying to production.
Join our Discord channel or contact us to get help or share feedback.
As the AI landscape is evolving, we are working to ensure Coder remains a secure platform for running AI agents just as it is for other cloud development environments.
Use Trusted Models
Most agents can be configured to either use a local LLM (e.g. llama3), an agent proxy (e.g. OpenRouter), or a Cloud-Provided LLM (e.g. AWS Bedrock). Research which models you are comfortable with and configure your Coder templates to use those.
Set up Firewalls and Proxies
Many enterprises run Coder workspaces behind a firewall or a proxy to prevent threats or bad actors. These same protections can be used to ensure AI agents do not access or upload sensitive information.
Separate API keys and scopes for agents
Many agents require API keys to access external services. It is recommended to create a separate API key for your agent with the minimum permissions required. This will likely involve editing your template for Agents to set different scopes or tokens from the standard one.
Additional guidance and tooling is coming in future releases of Coder.
Set Up Agent Boundaries Premium
Agent Boundaries add an additional layer and isolation of security between the agent and the rest of the environment inside of your Coder workspace, allowing humans to have more privileges and access compared to agents inside the same workspace.
Trial agent boundaries in your workspaces by following the instructions in the boundary-releases repository.