LLMs show promise for automated code analysis, but fall short in auditing real-world repositories due to context limitations and hallucinations. I present RepoAudit, an autonomous LLM-driven agent designed to perform precise, repository-level code auditing with high efficiency and accuracy. RepoAudit mimics expert auditors through demand-driven, path-sensitive reasoning over control and data-flow graphs—enabled by abstraction, pointer tracking, and validation mechanisms. Tested on 15 real-world projects in a controlled experiment, RepoAudit detected 38 true bugs with 65% precision, outperforming tools like Meta INFER and Amazon CodeGuru while costing only $2.54 per audit. A wider field-test has found 300 various kinds of zero-day bugs, ranging from classic bugs such as null pointer dereferences to functional bugs, in high-profile Github codebases, including Linux Kernel. This work represents a major step toward IDE-time, LLM-based auditing of large-scale software systems.