TokenFuzz is an open-source harness for evidence-driven, LLM-assisted security auditing. It runs coding agents against a source tree you are authorised to test. The agents read code, record concrete hypotheses, run testcases, and file what they find. The harness checks that evidence, and what survives review becomes a report a maintainer can act on.

It works with C/C++, Rust, Go, Swift, Python, Java, and other languages, from native libraries and command-line tools to browsers and JavaScript runtimes.

A prompt alone does not make a long audit reliable. The harness adds the parts that do:

• Source-to-testcase investigation. A deterministic ranker builds a shared work queue, and eight review strategies steer the analysis. No known bug or crashing seed is needed.
• Evidence-gated results. Every testcase runs through one probe command. A sanitizer crash must reproduce before it is promoted, and a concrete security issue that never crashes is still filed as a finding.
• Fleet coordination. Work leases, structured state, and clustering let parallel agents pick up each other's work instead of rediscovering the same root cause.
• Reviewable triage. Independent validation, reachability and caller-control fields, indexes of rejected results, and severity annotation make each model claim traceable rather than self-certifying.
• Measured review coverage. The run records every auditable file, agents attest the lines they read, their transcripts cross-check those receipts, and a report lists what the run never looked at. A clean result is not mistaken for a complete one.
• Maintainer handoff. Each accepted crash becomes a self-contained bundle: a report, the input, the sanitizer output, and a one-command script that reproduces it on a clean checkout.
• Comparable evaluation. A built-in benchmark runs TokenFuzz and a direct vulnerability prompt with the same target, model, and wall-clock budget, then compares validated, deduplicated evidence rather than volume of prose. See an example result page from a run against the C sample.

TokenFuzz drives Claude Code, Codex CLI, Gemini (through the Antigravity CLI or Google Gemini CLI), Grok Build, and OpenCode with either a catalog provider or a local OpenAI-compatible endpoint. Claude Code and Codex CLI run inside their own OS sandbox by default; Gemini, Grok, and OpenCode need a container or VM you administer. --backend all rotates the installed hosted backends that the chosen isolation mode can launch. Every backend runs the same audit contract, and bounded source reads, prompt reuse, execution budgets, and resumable state keep long runs manageable. Final security judgment stays with you and the upstream maintainer.

Quick start

git clone https://github.com/tokenfuzz/tokenfuzz
cd tokenfuzz
bash tests/run-tests.sh

# A one-worker smoke test against a shipped synthetic target.
bin/audit --target samples/sample-python --backend <backend> 1

Prerequisites lists the host tools and backend CLIs. Add a target and First audit walk through auditing your own project.

Documentation

The handbook covers installation, a first audit, target configuration, result triage, architecture, benchmarking, and development.

Only test software you are authorised to assess, and report target findings through the upstream project's security process.