As AI helps software developers generate code faster and automate tasks, it can also introduce security flaws that students may not recognize. Nishant Vishwamitra, PhD, assistant professor in the College of AI, Cyber and Computing, hopes to change that.
Vishwamitra is leading the $500,000 National Science Foundation–funded project to train students, before they enter the workforce, to spot security risks in AI-generated code.
OpenSecCoder, a free online learning platform developed through the project, will teach students to work with AI-generated software code and practice finding and fixing security problems in a safe environment. The project also will develop eight hands-on labs, three course modules and annual workshops for faculty.
“Most software today is written with AI help,” Vishwamitra said. “And those tools frequently produce insecure code. The problem is that students are not being taught to spot or fix those flaws.”
Vishwamitra added that no open platform, lab or curriculum currently exists for this type of training.
The project envisions an online training interface where students can watch attacks unfold and practice defending against them. The platform will offer a sandboxed environment that allows learners to safely observe, replicate, and experiment with different attack and defense techniques.
Vishwamitra described what a training lab would look like:
Students begin by selecting a topic and choosing the corresponding lab. This opens an interface that includes step‑by‑step instructions; a coding agent, which serves as the AI tool; a terminal for executing commands; and a debugger for identifying vulnerabilities. As students work through more labs, they gain experience using AI to support secure software development.
The project is expected to reach about 600 undergraduate and 400 graduate students each year in collaboration with the University at Buffalo.
Students will have the opportunity to participate in lab work, capstone projects, independent studies and summer workshops. Several labs will be designed for students without programming experience because AI tools are increasingly used in fields such as business, health care and communications.
Questioning AI-generated code
The program focuses in part on agentic AI. Unlike traditional AI chatbots that respond to individual requests from a user, agentic AI can take a larger task and carry out many steps on its own. It can create files, install software, run commands and correct its own mistakes without a person directing every action. AI coding tools such as Claude Code, Cursor and GitHub Copilot offer agentic capabilities that can perform multistep coding tasks with varying degrees of autonomy.
“Agents are built to finish the task and security is usually what gets sacrificed,” Vishwamitra said, adding that the code could end up in health care, finance, critical infrastructure, and defense systems.
Researchers want students from different majors to learn how to evaluate AI-generated code, test it and recognize possible security threats rather than automatically assuming the results are safe. Five of the eight labs will require no programming background, opening the training to students outside traditional computing fields.
“We want students to come away reading AI-generated code critically instead of trusting it, to be comfortable with professional scanning tools, able to measure and compare AI tools rather than picking one based on reputation,” Vishwamitra said, and noted that the goal is for students to be able to recognize an attack aimed at an agent.
Preparing a broader AI-ready workforce
The project is designed to help close the gap between rapidly evolving AI technology and the skills taught in the classroom. Training environments for the students will be modeled on real‑world professional settings and the project’s materials will be updated as new security threats emerge, Vishwamitra said.
The resources will also be available under an open-source MIT license, allowing community colleges, minority-serving institutions and rural universities to adopt and use them.
Vishwamitra sees secure AI use as a skill that will matter well beyond cybersecurity professions.
“Ultimately, we’d like security to stop being a specialist’s job,” he said. “If workers across every field can tell when an AI tool has handed them something unsafe, that changes the national security picture in a real way.”
