When Attackers Tune In: Weaponizing LLM Tuning for Stealthy C2 and Exfiltration

The talk is framed as a set of "what if" scenarios about how an attacker could disrupt or abuse a Kubernetes cluster, drawn from the open-ended questions the speakers ask job candidates. It opens with a primer on core objects such as pods, deployments, namespaces, service accounts, role bindings and secrets, and points out that a namespace is a soft security boundary that depends entirely on RBAC.

The middle section covers attacks. Patching the CoreDNS DaemonSet and scaling its replica set to zero breaks cluster name resolution, but the speakers note that the kubelet resolves image pulls through the node's own resolver, so some workloads keep running while others fail in cascading timeouts. A malicious validating admission webhook can quietly receive copies of new secrets, config maps and pods, and a live demo showed this with a small listening server. The speakers also show that a monitoring tool's exposed Prometheus-style API can reveal nodes, pods, containers, namespaces and load balancers in a single request that blends in with normal traffic.

For defenders, the takeaways are tighter RBAC, reviewing webhook configurations and alerting when they change, applying network policies around workloads, and checking whether a security product really sees both control-plane and runtime activity. The speakers stress that vendor visibility claims should be verified rather than trusted.