Can You Trust Your AI? Backdoors, Poisoned Prompts and Hacks That Cost Pennies - AgileLnL - Mark Shead
150 people attending
AI changes security in ways most teams haven't caught up with yet. In this session I'll give you a survey of the new things you need to think about, using real incidents and research from the last two years. We'll look at whether you can trust the model itself: how hidden behavior can be trained into a model where no one can see it, and what researchers have already found in real models. We'll look at poisoned prompts: why an email, web page or document can give your AI orders, and what happened when attackers tried it on Microsoft Copilot and Amazon's coding assistant. We'll talk about why the small "nobody would bother" bugs are now worth attacking when hacks cost pennies, and what AI-written code adds to the pile. I'll also share some of the ways we lower the risk, including heavy testing and having AI models from different labs check each other's work. No deep AI background needed. If you write software, manage people who do, or help decide which AI tools your team can use, you'll come away with a solid foundation for thinking about AI security.
