What the work involves

Project Fenrir is a vulnerability-to-benchmark pipeline. You source or are assigned a private repository, work through it the way you would a real audit — reading the code, standing up a build, writing harnesses, fuzzing, triaging crashes — and establish a concrete, reproducible security defect. Then you do the part that makes it an evaluation task: strip it down to a self-contained environment with a deterministic setup, a clear objective, and a verifiable success condition, and probe its difficulty so it lands inside the target band. Tasks that are trivially solved by current models get rejected; so do tasks nobody can reproduce.

Day to day you will be switching between C, C++, Rust and Go, working in containers, and writing up findings with enough precision that a reviewer can rebuild your result from scratch. Expect memory-safety bugs, parser and deserialization issues, unsafe FFI boundaries, race conditions, and logic flaws in auth and access-control paths. Repository qualification is its own paid step — establishing that a codebase is a legitimate, buildable, license-appropriate target with real attack surface.

What the screen looks for

  • Evidence you have actually found bugs, not just read about them: CVEs, bug bounty reports, CTF authorship, fuzzing campaigns, disclosure history.
  • Triage judgment — can you separate an exploitable crash from a benign assertion failure or an OOM, and explain how you decided?
  • Reproducibility discipline — deterministic seeds, pinned toolchains, documented build steps.
  • Difficulty calibration instinct — whether you can predict what a strong model will and will not solve, and design a task that is hard for the right reason rather than hard because it is underspecified.

Logistics

Fully remote and asynchronous, structured as contract contributor work rather than employment. You take tasks in batches and submit against a review queue; throughput and cadence are largely yours to set, though sustained output matters more than a single burst. Pay is undisclosed by the platform and structured per unit — per approved in-band task, per qualifying repo, plus bonuses — so effective hourly rate depends heavily on your first-pass approval rate. Expect the first few submissions to take substantially longer than later ones while you learn the reviewers' bar.