# The Human's Guide to Understanding AIxCC Source: https://www.youtube.com/watch?v=NTnw9Bm0l8w Recap page: https://rapidrecap.app/video/NTnw9Bm0l8w Generated: 2025-09-26T16:02:38.716+00:00 --- ## Quick Overview The AIxCC competition, focused on AI systems and bug hunting, is presented as a game that requires players to understand and interact with complex, non-visible systems, similar to how one might understand a sport through observation and analysis rather than direct participation. **Key Points:** - AIxCC is a competition focused on AI systems, requiring participants to find and fix vulnerabilities in code. - The game is inherently non-visible, making it difficult to understand without explicit visualization. - The presentation uses sports as a metaphor for understanding the game, highlighting visual observability, speed, scale, and human playability. - The competition involves multiple teams working on numerous challenges over several days, generating millions of LLM requests and telemetry actions. - Visualizations are key to understanding the game's complexity, showing the relationships between files and vulnerabilities. - Players score by identifying, reproducing, responding to, and patching vulnerabilities, mirroring the tasks of open-source maintainers. - The future of AIxCC could involve evolving into a more spectator-friendly sport with enhanced visualizations and narrative elements. ![Screenshot at 00:00: The presentation begins with the AIxCC logo, featuring a shield with circuit board motifs and the text "AI CYBER CHALLENGE," along with the tagline "AIxCC STAGE AT DEF CON 33."](https://ss.rapidrecap.app/screens/NTnw9Bm0l8w/00-00-00.png) **Context:** This presentation, titled 'The Human's Guide to Understanding AIxCC,' aims to demystify the AI Cyber Challenge (AIxCC) competition. The speaker, Mark Griffin, explains the nature of the competition, which involves finding and fixing vulnerabilities in AI systems. He uses the analogy of sports to make the abstract concepts more relatable, emphasizing the need for effective visualization to understand the complex, non-visible interactions within the AI systems being tested. ## Detailed Analysis The presentation "The Human's Guide to Understanding AIxCC" delves into the AIxCC competition, a cybersecurity challenge focused on finding and fixing vulnerabilities in AI systems. The core challenge lies in the fact that the "game" is not directly visible, making it difficult for humans to grasp without effective visualization. The speaker, Mark Griffin, draws parallels between AIxCC and traditional sports, highlighting aspects like visual observability, speed, scale, and the fundamental ease of understanding sports due to human participation. He notes that AIxCC involves multiple independent teams working on numerous challenges over extended periods, generating vast amounts of data (millions of LLM requests and telemetry actions). The presentation showcases the "field view" visualization, which maps out the relationships between code files and vulnerabilities, allowing observers to understand the game's dynamics. The scoring system mirrors open-source maintenance tasks: proving a vulnerability, responding to bug reports, and patching. The speaker also touches upon the difficulties in showing AIxCC, as every action is non-visible and requires explicit visualization. He concludes by suggesting that AIxCC, with its complex yet understandable system interactions and the potential for compelling narratives around teams and their progress, could evolve into a recognized "sport" with the right visual aids and commentary. ### Introduction - The AIxCC competition is presented as a game focused on AI systems, with the core challenge being its non-visible nature. ### Sports Metaphor - Human sports are used as an analogy to explain AIxCC, emphasizing visual observability, scale, speed, and understandability. ### Competition Scale - The final round involved 7 teams, 55 challenges, nearly 7 days of competition, approximately 2 million LLM requests, and tens of millions of telemetry actions. ### Visualizing the Game - A "field view" visualization was developed to represent the game's state, showing time, competitors, challenges, LLM activity, and scoring. ### Gameplay Observation - Visuals depict LLM resources and the flow of requests, with different colors representing different file types in the repository structure. ### Scoring Mechanism - Teams score by identifying, reproducing, responding to, and patching vulnerabilities, similar to open-source maintenance tasks. ### Challenges - Repositories with added vulnerabilities, variety in tasks (C vs. Java, Delta vs. Full scans), synthetic vulnerabilities, and SARIF broadcasts are key challenges. ### Difficulty of Understanding - The non-visible nature of AI systems and the sheer scale of operations make it hard to intuitively grasp what's happening, requiring detailed visualizations and analysis. ![Screenshot at 00:00: The presentation begins with the AIxCC logo, featuring a shield with circuit board motifs and the text "AI CYBER CHALLENGE," along with the tagline "AIxCC STAGE AT DEF CON 33."](https://ss.rapidrecap.app/screens/NTnw9Bm0l8w/00-00-00.png) ![Screenshot at 00:04: A glitch effect overlays the AIxCC logo and background code, transitioning to the title "THE HUMAN'S GUIDE To Understanding AIxCC."](https://ss.rapidrecap.app/screens/NTnw9Bm0l8w/00-00-04.png) ![Screenshot at 00:16: The screen displays the title "AIxCC FINAL ROUND" and a timeline graphic representing 55 tasks over approximately 7 days, split into "Full Scan" \(red\) and "Delta Scan" \(blue\) blocks.](https://ss.rapidrecap.app/screens/NTnw9Bm0l8w/00-00-16.png) ![Screenshot at 01:17: The speaker, Mark Griffin, stands at a podium on stage with the AIxCC logo prominently displayed on a large screen behind him.](https://ss.rapidrecap.app/screens/NTnw9Bm0l8w/00-01-17.png) ![Screenshot at 01:27: A "Lord of the Rings" fellowship meme is displayed, with the fellowship members' faces replaced by question marks, symbolizing the unidentified finalists.](https://ss.rapidrecap.app/screens/NTnw9Bm0l8w/00-01-27.png) ![Screenshot at 01:31: The same "Lord of the Rings" meme is shown, but now with the logos of the 7 finalist teams superimposed on the fellowship members.](https://ss.rapidrecap.app/screens/NTnw9Bm0l8w/00-01-31.png) ![Screenshot at 04:09: A slide titled "New-to-a-sport Questions for AIxCC" lists key questions like "What's the goal?", "What are the teams?", "How long is the game?", and "What are the rules?"](https://ss.rapidrecap.app/screens/NTnw9Bm0l8w/00-04-09.png) ![Screenshot at 04:35: A slide titled "The Difficulties of Showing AIxCC" outlines challenges such as the competition being built for custom systems, non-visible actions, and the need for explicit visualization.](https://ss.rapidrecap.app/screens/NTnw9Bm0l8w/00-04-35.png) ![Screenshot at 05:46: A comparison of "Scale and Speed of Sport \(Soccer\)" with "Scale and Speed of AIxCC \(final round\)" highlights the vast difference in metrics, e.g., 22 players vs. 7 teams, ~1000 passes vs. ~2 million LLM requests.](https://ss.rapidrecap.app/screens/NTnw9Bm0l8w/00-05-46.png) ![Screenshot at 07:53: A slide titled "How do you score?" outlines the process mirroring open-source maintainers: proving a vulnerability, responding to a bug report, patching, and bundling submissions, with checkmark and cross icons representing success and failure/rejection respectively.](https://ss.rapidrecap.app/screens/NTnw9Bm0l8w/00-07-53.png)