eBPF enables transparent control of AI API traffic in Kubernetes without code changes

The presentation points out that AI-generated code often enters production without clear ownership, leaving teams unable to support or verify its behavior. This creates hidden dependencies that can cause incidents when the code fails or acts unexpectedly. Engineers must therefore find ways to observe and govern AI agents after they are deployed. The talk proposes using eBPF as a runtime tool to address this gap.

eBPF allows kernel-level socket hooks that can capture network traffic to and from AI services running in Kubernetes. By attaching these hooks, operators can inspect prompts, responses, and associated metadata without modifying the application containers. The demonstrated use cases include filtering prompts, swapping models, enforcing token limits, and restricting specific system calls. All of this is done transparently, meaning the AI agent sees no change in its interaction.

Adopting this approach requires familiarity with eBPF programming and the tooling provided by projects such as Cilium, which the presenter helps maintain. Teams must write and deploy eBPF programs that implement the desired security policies. There may be a performance cost associated with packet inspection, though the presenter notes that overhead is typically low for socket-level hooks. Additionally, the host nodes must run a recent Linux kernel that supports the needed eBPF features.

The technique works best in Kubernetes clusters where all nodes run Linux and the AI services expose traffic over standard sockets. Encrypted traffic that is not terminated before the eBPF hook may limit visibility into payloads unless decryption is performed upstream. The approach primarily governs API-level interactions and may not capture behaviors such as local file system calls or subprocess invocations that bypass the network layer. Consequently, engineers should complement eBPF monitoring with other observability and runtime security measures.