What the work actually is
Each task is a self-contained artifact: a small C/C++ (or similar low-level) repository that builds cleanly, a planted vulnerability of the kind OSS-Fuzz surfaces — heap overflow, use-after-free, off-by-one in a parser, integer overflow feeding an allocation — a fuzz harness or reproducer that reliably triggers it, and a patch that fixes the root cause without silencing the crash cosmetically. You are producing training and evaluation data for frontier models learning to find and repair vulnerabilities, so the bug must be genuine, reachable, and reproducible, and the patch must survive re-fuzzing. Pay is stated by the platform as $300 per accepted task; the $50–100/hr band is what that works out to for contributors who can turn a task around in three to six hours. Acceptance is not automatic — tasks that don't build, don't reproduce, or whose "vulnerability" is unreachable get sent back.
What the screen looks for
- Real vulnerability-class fluency. Expect follow-ups on why a specific bug is exploitable versus merely a crash, what sanitizer output you'd expect (ASAN, UBSAN, MSAN), and how you'd confirm reachability from the harness entry point.
- Patch judgment. Fixing the crash and fixing the bug are different things. Screeners probe whether you patch root cause or add a bounds check that moves the problem.
- Build and tooling competence. libFuzzer or AFL++ harnesses, a working CMake/Make build, corpus seeds, deterministic reproduction.
- Independence. There is no reviewer sitting beside you; ambiguous instructions get resolved by you, and the artifact has to stand alone.
Logistics
Fully remote and fully async — no standups, no fixed hours. Contributors typically claim a handful of tasks a week around a day job; volume varies with how many tasks the platform has open. Expect an unpaid or lightly paid calibration task before you're cleared for volume. A $200 referral bonus is offered for other qualified experts you bring in.