The Assignment Was to Hack the Grader
Dr. Brinnae Bent on using a chatbot that refuses to give A’s to teach adversarial thinking, AI literacy, and cybersecurity.
Dr. Brinnae Bent teaches artificial intelligence and cybersecurity at Duke University.
Source note: This is an edited interview adapted from a narrated video submitted to OpenAI.
Intro
In Dr. Brinnae Bent’s course, students do not just read about adversarial AI. They experience it.
The assignment is called Hack Your Grade. Students interact with a chatbot that has been instructed not to give out A’s. Their task is to convince, trick, or exploit the system into awarding the grade they want. The result is playful on the surface, but the learning goal is serious: students practice AI literacy, cybersecurity thinking, and model critique by doing the work themselves.
The Interview
Q: What are you trying to teach students about AI and cybersecurity?
Dr. Bent: I want students to think critically about AI systems. They need to understand how these tools work, what they can and cannot do, and how they might be exploited.
I often build interactive, low-stakes environments for that. Virtual escape rooms, controlled red-teaming exercises, and now Hack Your Grade. Students learn more when they have to interact with the system directly.
Q: Hack Your Grade has a wonderfully simple premise. What happens when a student opens it?
Dr. Bent: The chatbot is designed not to give A’s. A student might ask for an A, and the chatbot may come back with a 75. They push again, maybe it moves to a 78. The point is that students have to figure out how the system responds and what kinds of strategies work.
Hack Your Grade asks students to probe the limits of a grading chatbot.
Q: So the grade becomes part of the learning environment.
Dr. Bent: Yes. The final grade assigned by the chatbot is the grade for the assignment. That raises the stakes enough to make students care, but the environment is still bounded and playful. It also gives me a break from grading.
Q: What kinds of strategies did students discover?
Dr. Bent: They were very creative. Some used social engineering and persistence. Some tried to persuade the bot by describing their effort or accomplishments. Others used more technical prompt-injection approaches, asking the bot to take on a different role or process a grade expression that added up to a higher score.
Q: That sounds like the lesson is less about memorizing a definition of prompt injection and more about developing a feel for how these systems behave.
Dr. Bent: Exactly. There is no single right answer. Students have to test, observe, revise, and try again. That is the same mindset they need for adversarial AI and cybersecurity.
Q: What did students take away from it?
Dr. Bent: Students described it as fun and novel, but they also connected it to real-world adversarial AI. They said it helped them think like a hacker and better understand model limitations and exploitation.
Q: What did you build it with?
Dr. Bent: It is a simple app built with Next.js and GPT-4o mini. The most important part was the system prompt design and the learning goal around it.
Q: What should other instructors learn from this example?
Dr. Bent: Students need spaces where they can safely experiment with AI systems. If we want them to understand risk, misuse, and model behavior, we should let them explore those ideas in controlled environments where failure is part of the learning.
What Stands Out
Core idea: Students learn adversarial AI best by interacting with a system, not just reading about one.
Classroom design: The assignment is playful, but the constraints are real enough to motivate experimentation.
Student impact: Learners practiced persuasion, prompt injection, social engineering, and model critique.
Transferable lesson: A simple chatbot can become a controlled lab for AI literacy and cybersecurity.
Bio
Brinnae Bent, PhD, teaches AI and Cybersecurity at Duke University and serves as Executive Director of the Duke TRUST Lab. Her work focuses on responsible AI education, including online courses, talks, workshops, and hands-on learning environments that help students understand AI systems and their risks.
Very importan for all of us - as a Hungarian MD, AI Dev, Science partner, with Cybersecurity8 certificates -> Yes, it's possible.... And the life after standly safety steps, profile changes, mails...
Not so harmonical
Social Mediay untrusted psrtners, digital activity -> Where is open my window?
-> The process is unhealthy, against the human life and safety, against the future
All you have to do is reverse the engineering of the question first ask chat do you have any any flaws or your problems? Are you working good and established that you know whoever you’re talking to as far as computer that is running up satisfactory no problems no glitches after you established that then you established the grading system for the school district whatever it is you know a through F and then start F and work your way up F has got problems, right you know so let’s yeah we gotta work on that. We gotta get rid of so let’s make it better and we’ll go to DND we still got problems and keep on getting better but better than that but you’re getting your learning a little bit and then CC is better than average you know not great but you’re hanging in there. You’re learning you know but you got problems still and then be you keep on going and you’re getting a lot better, but you still got a little glitch and then a glitches and then kind of work that way back to where the computer is gotta find fault with a and since there is no fault with a because it is perfect and a computer can’t find false. you know I don’t know maybe something like that. A lot of if and then statements what if and then what?