Harmful biases in large language models (LLMs) make AI less trustworthy and secure. Auditing for biases can help identify potential solutions and develop better guardrails to make AI safer. In this podcast from the Carnegie Mellon University Software Engineering Institute (SEI), Katie Robinson and Violet Turri, researchers in the SEI’s AI Division, discuss their recent work using role-playing game scenarios to identify biases in LLMs.

In the wake of widespread adoption of artificial intelligence (AI) in critical infrastructure, education, government, and national security entities, adversaries are working to disrupt these systems and attack AI-enabled assets. With nearly four decades in vulnerability management, the Carnegie Mellon University Software Engineering Institute (SEI) recognized a need to create an entity that would identify, research, and identify mitigation strategies for AI vulnerabilities to protect national assets against traditional cybersecurity, adversarial machine learning, and joint cyber-AI attacks. In this SEI podcast, Lauren McIlvenny, director of threat analysis in the SEI’s CERT Division, discusses best practices and lessons learned in standing up an AI Security Incident Response Team (AISIRT).

The exposed and public nature of application programming interfaces (APIs) come with risks including the increased network attack surface. Zero trust principles are helpful for mitigating these risks and making APIs more secure. In this podcast from the Carnegie Mellon University Software Engineering Institute (SEI), McKinley Sconiers-Hasan, a solutions engineer in the SEI CERT Division, discusses three API risks and how to address them through the lens of zero trust.