HarnessLLM preprint derives Rust verification harnesses from test suites and reports six memory-safety bugs

Sources: arXiv 2607.22161
Summary: A preprint describes HarnessLLM, which derives Rust verification harnesses from existing test suites. It reports extracting 294 calling scenarios from 494 test cases at 94.66 percent precision, then generating harnesses for all of those scenarios, while Autoharness succeeded on only 41 percent of them. Those are two different measurements: 94.66 percent is scenario-extraction precision, and the figure comparable to Autoharness's 41 percent is coverage of all scenarios. It reports six real memory-safety bugs found. Kani appears nowhere in the abstract or metadata read at this run, though Autoharness is a Kani tool. This is a single preprint, the figures are the authors' own, and the six bugs are not identified here.
Why it matters: Harness authoring is the step that keeps bounded model checking out of most Rust projects, and finding real bugs rather than reporting a benchmark delta is the result that matters for anyone deciding whether to wire a model checker into CI.