AI Agent Security Has to Be Layered
Blog post from Coder
AI coding agents inherit the credentials, network access and permissions of their environments, so prompts and in-process approval controls cannot replace security boundaries enforced outside the agent. A reference design on Amazon EKS using AWS Nitro Sandboxes illustrates a layered approach involving isolated workspaces, admission and network policies, per-workspace host allowlists, shared egress controls, scoped cloud identities and centralized AI governance. Testing a second agent, Kiro CLI, showed that it could bypass the AI Gateway by using its own service, while infrastructure controls still constrained its network destinations, AWS access and runtime; therefore, coverage must be assessed for each agent and traffic path rather than assumed from an architecture diagram. The design also depends on consistent orchestration, separation of duties, automated tests and a fast, secure workspace setup, since configuration gaps can undermine multiple controls. Host allowlists limit destinations but cannot distinguish repositories or judge whether an authorized action is appropriate, so stronger measures such as a Git broker or narrowly scoped TLS inspection may be needed for specific risks, alongside branch protections, reviews and deployment approvals. The recommended approach is to isolate and constrain workspaces, route model traffic through approved governance where possible, scope credentials and storage, bound and monitor egress, secure the surrounding platform, and add more intrusive controls only when risk warrants them.
| Trend | Post Mentions | Total Month Mentions | Posts | Companies | MoM |
|---|---|---|---|---|---|
| MCP | 2 | 9,757 | 959 | 203 | -18% |
| AI Agents | 1 | 6,233 | 1,230 | 245 | -11% |
| AI Coding Assistant | 1 | 1,849 | 451 | 145 | -6% |
| Kubernetes | 1 | 4,377 | 419 | 114 | -15% |
| Platform Engineering | 1 | 1,295 | 290 | 84 | -36% |
Use this post, company, and trend context to find content marketing opportunities, perform competitive analysis, or address product feature gaps via the Plushcap MCP server or the Plushcap API.