Skip to content
Book Open access

LDBF: Deep Exploration of eBPF Runtime via LLM-Driven Program Generation and Adaptive Context Orchestration

Nov 2026 · Proceedings of the 2026 ACM SIGSAC Conference on Computer and Communications Security · 0 citations · 22 references

Abstract

Extended Berkeley Packet Filter (eBPF) substantially enlarges the Linux kernel attack surface by executing user-supplied programs in kernel space. While prior efforts have intensively studied the static eBPF verifier, the dynamic eBPF runtime remains insufficiently tested. Fuzzing the eBPF runtime is challenging for two reasons: (i) generating diverse programs that satisfy the verifier's strict semantic and dependency constraints, and (ii) constructing compatible yet diverse execution contexts (e.g., attach points) so that the same verifier-accepted program can be executed under multiple contexts to trigger different runtime behaviors. Existing kernel fuzzers and eBPF-specific fuzzers struggle to achieve both of them. We present LDBF, a runtime-oriented fuzzing framework that decouples probabilistic eBPF program synthesis from deterministic context orchestration. First, LDBF leverages extracted BPF domain knowledge and an LLM-guided synthesis pipeline with a compiler–verifier feedback loop to produce diverse, verifier-accepted eBPF programs. Second, it introduces an adaptive orchestration mechanism that infers each seed program's required resources and deterministically synthesizes the corresponding runtime environment to maximize triggerability and runtime coverage. We evaluate LDBF on recent Linux kernels. Compared to state-of-the-art eBPF runtime fuzzers, LDBF increases eBPF component coverage by 27.3% on average (23.2%–30.4% over individual baselines) and discovers more than 30 new vulnerabilities.

Read PDF

We use cookies to run the site and, with your consent, for analytics and to show ads. See our Cookie Policy.