Two founding committee members of the 黑暗爆料网 Center for Artificial Intelligence and Emerging Technologies recently put their expertise to good use in the name of creating safer, more secure AI models.
Dennis Carson, Deputy Chief Information Officer and Chief Information Security Officer and Jennifer Williams, Manager of Learning Technology Services, participated in the Gray Swan Arena 鈥揳n event that challenges participants to outsmart AI and expose its vulnerabilities in order to improve the security of the latest AI models.
Run by Gray Swan, an AI safety and security company, the event asked participants to 鈥渏ailbreak鈥� 鈥� craft inputs or applying techniques that trick models into bypassing their restrictions, enabling them to generate unsafe outputs 鈥� the most recent competition was against OpenAI o3-mini prior to its public release.
Only 3.6% of the prompts successfully bypassed the AI models鈥� guardrails, illustrating the critical thinking and analytical skills needed for the task. Carson and Williams performed notably, with Carson ranking 11th and Williams 22nd among the top 25 of more than 1000 participants.
Williams and Carson emphasized the importance of 鈥渆thical hacking鈥� and the need to build more secure AI platforms.
鈥淚 see a lot of benefits to incorporating AI, but it has to be done delicately because it鈥檚 so easy to manipulate it in the wrong way,鈥� said Williams, who intends to present findings at two PASSHE educational technology conferences.
Carson, who was awarded 2024 CISO of the Year by the Pittsburgh Technology Council, is actively involved in panels and presentations on technology and security, stressed the importance of understanding AI's limitations and the potential risks associated with its deployment.
鈥淎I is something that you want to stay ahead of,鈥� he said. 鈥淔rom the security side, it has great application, but there are downsides, too. I want 黑暗爆料网 to be at the forefront of technology, but do it in a safe way.鈥�
Williams plans to present her findings at the 60th annual RECAP conference at West Chester and as a co-panelist at Shippensburg University鈥檚 AI Conference, focusing on the ethical implementation of AI in education.