Poster 10 (in-person)
Poster 10 (in-person)
Embedded Files
Tessa Hutchman, Sanchali Sharma and Ozgur Ozkan, audn.AI
Tessa Hutchman, Sanchali Sharma and Ozgur Ozkan, audn.AI
Co-founders
Co-founders
Tessa Hutchman the co-founder of a company called Audn.ai, a cybersecurity company that works with LLMs.
Tessa Hutchman the co-founder of a company called Audn.ai, a cybersecurity company that works with LLMs.
Sanchali Sharma is the CPO & co-founder building secure, production-ready AI agent systems that bridge frontier AI research and real-world deployment; ex-Microsoft, shipping large-scale AI products and zero-to-one ventures.
Sanchali Sharma is the CPO & co-founder building secure, production-ready AI agent systems that bridge frontier AI research and real-world deployment; ex-Microsoft, shipping large-scale AI products and zero-to-one ventures.
Ozgur Ozkan is a co-founder of audn.ai, where he builds automated adversarial testing for LLM-powered voice and language systems. His MSc thesis applied representational similarity clustering to large-scale face recognition, and his published work includes the use of machine learning to identify novel SARS-CoV-2 inhibitors. He holds six patents in video and multimedia technology and has trained ML models in production across chemistry, medical imaging, and generative video.
Ozgur Ozkan is a co-founder of audn.ai, where he builds automated adversarial testing for LLM-powered voice and language systems. His MSc thesis applied representational similarity clustering to large-scale face recognition, and his published work includes the use of machine learning to identify novel SARS-CoV-2 inhibitors. He holds six patents in video and multimedia technology and has trained ML models in production across chemistry, medical imaging, and generative video.
Email: tessa@audn.ai
AudnAI_Poster_A1_Print_5.pdfTitle: "Adversarial Risks in AI-Powered Language Assessment Platforms"
Authors: Ozgur Ozkan, Arun Baby, Sanchali Sharma, Tessa Hutchman
Abstract: As Large Language Models are deployed in high-stakes language assessments—scoring essays, conducting oral examinations, and generating test items—they introduce a novel attack surface absent from traditional psychometric frameworks. This poster synthesises emerging empirical evidence on adversarial vulnerabilities in LLM-based assessment systems, mapping attack vectors (prompt injection, role-based jailbreaks, voice social engineering) to their downstream impact on fairness and validity. We show that simple, low-sophistication attacks reliably manipulate LLM scoring outcomes across multiple domains, and propose continuous adversarial validation as a prerequisite for responsible deployment. Human oversight alone is insufficient: automated red-teaming must become part of the assessment quality assurance lifecycle.
Page updated
Report abuse