Open-source security triage · NVIDIA Nemotron on Nebius Token Factory
AI made fake security reports free to produce.
ProofGate makes proving them cheap.
Every incoming vulnerability report is checked claim by claim against the real source code, at the exact version it names, before a volunteer maintainer spends an hour on it. Confirmed bugs go all the way to a tested patch. A human always makes the final call.
The problem
A fake report costs seconds to generate and an hour to disprove.
Open-source projects like curl run on billions of devices and are maintained by volunteers. AI-generated reports cite functions that never existed, line numbers past the end of the file, and code nobody wrote. The curl project keeps a public list of them. ProofGate turns that hour into seconds by checking the claims against git history, not by trusting another model's opinion.
Grounding, visualized
Every claim is checked against every commit since 1998.
When a report cites a function, ProofGate looks for it at the claimed release, on the branch as of the day the report was filed, and across the project's entire history with git's pickaxe. For curl that is the whole helix below, searched in about four seconds, at no model cost.
git log -S"Curl_cookie_parse_header("How it works
Nine stages. Cheap checks first, expensive reasoning last.
- 01IntakeReads text, screenshots and terminal capturesNemotron Super VL
- 02ClaimsEvery checkable claim as JSONNemotron Nano
- 03Groundinggit + tree-sitter at the claimed versionNemotron Super
- 04Prior artAdvisories, disclosures, duplicatesTavily + OSV
- 05PolicyThe project's own SECURITY rulesNemotron Super
- 06ReproduceReporter's steps, no networkOpenShell sandbox
- 07VerdictVerdict + adversarial challengeNemotron Ultra
- 08FixPatch, rebuild, test, draft PRNemotron Super
- 09EvidenceLine-linked evidence card
Built on
Every model call, through the OpenAI-compatible API. Four Nemotron sizes, each with one job, with live cost accounting and a hard per-report budget.
Super VL reads screenshots, Nano extracts claims, Super reasons over real source and writes patches, Ultra decides and then argues the other side.
Runtime prior-art search across disclosed reports, advisories, issues and mailing lists, to catch duplicates and already-fixed bugs.
The reporter's own reproduction runs in an ephemeral sandbox with network denied, no credentials, and resource limits.