1 in 4 AI patches that make a crash go away isn't a fix.

Leakproof repairs memory bugs in C, C++ and CUDA, then tries to break every patch. It ships only the ones that survive.

Re-checking runs the full gate again on a real patch from our runs. No AI, no cost, about 30 seconds.

Patches that silenced a real OSS-Fuzz crash

Loading the run records…

Four NVIDIA Nemotron models on Nebius Token Factory, each with one job. The gate runs separately, offline, and decides.

ReproduceDocker or Nebius SandboxesBuild the project with sanitizers and confirm the crash.
DiagnoseNemotron 3 UltraFind the root cause and propose up to four fixes.
Patch4 × Nemotron 3.5 LightningEach writes one patch on its own copy. None sees the tests.
AttackNemotron 3 SuperWrites inputs meant to break each surviving patch.
GateSanitizers, tests, re-fuzzingShip with the evidence, or refuse and say why.

Reported in October 2026, after the models' training data. No fix existed to copy.

Project	Bug	Leakproof	Our review

A patch for a real jq bug passed jq's own tests. Leakproof doubled one number in the crash input and the patched jq overflowed.

Pick a run and a bug to see each patch, each check and its evidence.

Run

Pick a bug from the list.

Does the gate accept correct fixes?

CUDA on an NVIDIA T4

Model comparison

All runs

Run	Bugs	Crash gone	Verified	Model $ per verified fix

Limits

The old OSS-Fuzz bugs were fixed upstream years ago, so the models may have seen those fixes in training. The still-open bugs above are the clean test.
The gate proves the reported crash is gone and that nothing it tried causes a new sanitizer report. It can't prove a fix correct for every input, which is why a person reviews each fix before it is sent.
Without project tests, a verified patch is labelled "verified, needs human review".
Leak results come from Juliet, which is synthetic; the OSS-Fuzz set has no leak cases.
Nebius Sandboxes runs Juliet and bring-your-own projects; the OSS-Fuzz images run on local Docker.