DebugHarness: Emulating Human Dynamic Debugging for Autonomous Program Repair

A research team from Nanjing University has unveiled DebugHarness, a novel AI agent designed to autonomously repair complex, low-level security vulnerabilities like use-after-free and memory corruption in C/C++ software. The system addresses a critical bottleneck in software engineering: while fuzzers can find bugs, fixing deep-rooted, memory-related flaws still demands expert manual analysis. Current LLM-based repair tools treat bug fixing as a static code-generation task, missing the crucial dynamic execution context needed for accurate diagnosis.

DebugHarness overcomes this by emulating the interactive, investigative workflow of a human systems engineer. When triggered by a reproducible crash, the agent doesn't just look at static code. Instead, it formulates hypotheses and then actively probes the live program's memory states and execution paths. It operates in a closed-loop validation cycle, synthesizing patches and testing them against the runtime environment. This dynamic debugging approach proved dramatically more effective than static methods.

Evaluated on SEC-bench, a rigorous dataset of real-world C/C++ security bugs, DebugHarness successfully patched approximately 90% of the vulnerabilities. This performance represents a relative improvement of over 30% compared to state-of-the-art baseline models. The results demonstrate that integrating dynamic, context-aware investigation significantly enhances an LLM's diagnostic and repair capabilities for intricate systems-level problems.

The research establishes a new paradigm for automated program repair, effectively bridging the gap between the static reasoning of large language models and the dynamic, state-dependent intricacies of low-level systems programming. It moves AI beyond simple code generation and into the realm of interactive, environment-aware problem-solving, promising to reduce the manual burden on security engineers and accelerate the patching of critical software flaws.