Today’s advanced cyber attack campaigns can often bypass all existing protections. The primary defense against them is after-the-fact detection, followed by a forensic analysis to understand their impact. Such an analysis requires audit logs (also called provenance logs) that faithfully capture all activities and data flows on each host. While the Linux auditing daemon (auditd) and sysdig are the most popular tools for audit data collection, a number of other systems, authored by researchers and practitioners, are also available. Through a motivating experimental study, we show that these systems impose high overheads, slowing workloads by 2× to 8×; lose a majority of events under sustained workloads; and are vulnerable to log tampering that erases log entries before they are committed to persistent storage. We present a new approach that overcomes these challenges. By relying on the extended Berkeley Packet Filter (eBPF) framework built into recent Linux versions, we avoid changes to the kernel code, and hence our data collector works out of the box on most Linux distributions. We present new design, tuning and optimization techniques that enables our system to sustain workloads that are an order of magnitude more intense than those causing major data loss with existing systems. Moreover, our system incurs only a fraction of the overhead of previous systems, while considerably reducing data volumes, and shrinking the log tampering window by ~ 100×.