Scientist, Centre for Frontier AI Research, IAIC

The Trustworthy AI team at the Centre for Frontier AI Research (CFAR), IAIC, is seeking motivated scientists to conduct research on the safety of generative AI (GenAI) and multi-agent systems. Working with regulatory and clinical partners, the team develops the scientific evidence, testing tools and guidance needed for the safe and responsible deployment of AI solutions in real-world.GenAI and multi-agent systems, such as clinical chatbots built on large language models (LLMs) and vision-language models (VLMs) that interpret medical images, introduce risks that existing regulatory frameworks were not designed to address. These include open-ended outputs, hallucinations, the lack of accepted safety metrics, and vulnerability to jailbreaks, prompt injection and data leakage. You will conduct cutting-edge R&D to assess, test and mitigate these risks, and translate your findings into regulatory-grade toolkits and guidance with real-world meaningful impact.Key responsibilitiesDesign and conduct systematic red-teaming and risk assessments of GenAI and multi-agent systems, covering adversarial attacks, hallucinations, privacy leakage and performance across diverse scenariosDevelop efficient guardrails and risk-mitigation methods, such as knowledge validation, hallucination and uncertainty detection, adversarial input detection, and automated flagging of high-risk outputs for human reviewDevelop methods for continuous post-deployment monitoring, including near real-time performance tracking, drift detection and anomaly alertingCompetencies1. Integrity2. Critical Thinking3. Innovativeness4. Effective CommunicationAchieving OutcomesIDARE Behaviours - 4. Reach Outward, Tackle Real-World ChallengesResearch areas (one or more of the following)Red-teaming and adversarial robustness of LLMs, VLMs and LLM-based agents, e.g., jailbreaks, prompt injection, data poisoning, backdoors and model extractionHallucination detection and mitigation, uncertainty quantification and calibration for medical GenAIGuardrails and runtime safety mechanisms for clinical conversational agents and multimodal assistantsRobustness to distribution shift, and fairness across patient populations, clinical sites and imaging devicesRequirementsPhD degree in Computer Science, Electrical & Electronic Engineering, Computer Engineering, Biomedical Engineering or a related fieldSolid programming skills in Python and hands-on experience with deep learning frameworks such as PyTorchResearch experience in one or more of the following: AI safety and security, adversarial machine learning, trustworthy machine learning, LLMs or VLMs, and multi-agent systemsGood communication and teamwork skills, and the ability to work with clinicians, regulators and engineers in a multidisciplinary setting