Hacking for Defenders: approaches to DARPA’s AI Cyber Challenge

The report describes Google’s methodology for tackling DARPA’s AI Cyber Challenge exemplar by leveraging OSS-Fuzz and adapting AFL for kernel fuzzing with KCOV/KSAN instrumentation, a QEMU environment, and a userspace harness wrapper. It identifies challenges such as complex input formats, slow execution, and resource leaks, and proposes AI-driven enhancements—LLM-generated seed inputs, augmented static analysis, and LLM-based automated patching and validation—to accelerate finding and fixing vulnerabilities, with the goal of integrating advances back into OSS-Fuzz to benefit the broader open-source ecosystem.