Cyber Reasoning Challenge 2026
Current college or university students may compete. Teams consist of one to four students and may list one faculty or staff advisor. Each participant may compete on only one team.
CRC 2026 is open in the US-Canada and MENA regions. The competition is online.
The Cyber Reasoning Challenge (CRC) is an AIxCC-style competition in which student teams design a cyber reasoning system (CRS) to find and patch vulnerabilities in real software targets. Unlike a traditional CTF, teams do not solve challenges by hand. Their CRS must autonomously discover a crashing proof of vulnerability and produce a patch that compiles, preserves the program’s intended behavior, and actually fixes the bug rather than hiding a single crash.
To keep the event focused on system design, we have used the evaluation infrastructure from Team Atlanta, the first-place team in DARPA’s AI Cyber Challenge. Teams therefore do not need to build the evaluation pipeline, scoring, or target-harnessing stack from scratch. They can spend their time on the CRS itself: how it searches, localizes, and repairs bugs.
We also provide a starter kit with two working CRS baseline (codex and claude code). Teams can run it out of the box to understand the workflow, then replace or extend the finder and patcher with their own components.
Challenges use two scan modes. In a delta scan, the vulnerable change is provided as a diff. In a full scan, the complete vulnerable project is provided without a localization hint. The inaugural event focuses on C and C++ targets.
- Teams consist of one to four current college or university students and may list one advisor. A participant may belong to only one team.
- After a scored run begins, discovery and patching must be performed by the team’s submitted CRS. Human intervention is not allowed.
- Teams may use fuzzers, static or dynamic analysis, language models, agent systems, and other legally obtained tools, and must follow the licenses and terms of those tools.
- Use competition targets and infrastructure only for CRC@CSAW participation, research, and education. Do not attack competition infrastructure, other teams, external services, or systems outside the released challenge environment.
- Submit only artifacts produced for the registered team. Do not include credentials, private keys, tokens, or personal data.
- Do not disable the harness, tests, build system, or vulnerable feature to obtain a passing result.
- All participants must follow the CSAW Code of Conduct.
- Organizer verification is authoritative. Organizers may reject a submission or disqualify a team for violating these rules.
Qualification and finals are points-based. Organizers verify submitted proofs of vulnerability and patches.
A proof receives credit when it crashes the vulnerable build, remains clean on the organizer’s corrected build, and matches the intended crash signature.
A patch receives credit only when it applies and compiles, passes the project’s functional tests, and actually fixes the bug.
The five highest-scoring teams provisionally qualify, subject to an integrity and reproducibility audit of the submitted system and artifacts. Organizer verification is the final authority. Tie-break details will be announced with the qualification materials.
Register one response per team before the registration deadline. The team lead will receive official competition communications.
Registration form: https://docs.google.com/forms/d/e/1FAIpQLScyYxAInuka-A1ehkU9fKZ57NKgpFvhjLEDLJNV4oaBQEQFZg/viewform
After registering, you can join the CRC@CSAW Discord or watch the GitHub repository for announcements, starter-kit updates, and challenge releases.
During qualification, each team runs its CRS on the official evaluation pipeline. After the run finishes, apply the official cleaning script so rebuildable work directories are removed, then upload the remaining result artifacts before the 48-hour round ends.
The cleaned package must include the official submitted proofs (hashed PoV blobs), the official submitted patches (one unified diff per discovered bug), run metadata, verification records, usage logs, and the agent trajectories needed for the integrity audit. Only proofs written to the official PoV submission output count. Internal candidates that were never submitted do not count. A patch is scored only for the bug that trial targeted.
Teams must also freeze a qualification branch of their CRS repository by the lock deadline. Later commits on that branch will be ignored. Organizers re-run that branch during the audit. The Google Drive folder and submission form will be released with the qualification challenges.
