Security for AI Basics - Not by ChatGPT

The talk opens with a light icebreaker and a case for why AI security matters now: AI tools already sit inside everyday products, and many organizations do not have a clear inventory of the models they run. The speaker, who leads threat intelligence at a security firm, uses quick true-or-false questions, analogies and jokes to keep the session approachable for newcomers.

The middle section separates an AI model from the wider AI system around it, including data pipelines, interfaces and deployment tooling, and names the likely adversaries: nation states, competitors and cybercriminals. It then walks through the main attack classes. Data poisoning corrupts training or continuously learned data. Evasion uses adversarial inputs that look normal to people but cause misclassification, illustrated with a stop sign altered by stickers that fooled a self-driving car. Model theft uses repeated queries to rebuild a model's behavior. Prompt injection is covered in direct, indirect and multimodal forms, along with cases where a language model's generated code is executed on a server. Supply chain risks close the section, covering backdoored pre-trained models and unsafe serialization formats in widely used machine learning libraries.

The closing part looks at policy, contrasting the EU's restrictive AI act with the more market-driven US approach, and notes that AI red teaming standards are still being defined. It ends with lifecycle recommendations: check data sources and model provenance at design time, scan and verify model integrity before deployment, then monitor inputs and outputs, run adversarial testing and educate users afterward. The overall takeaway is that AI security needs proactive work across teams because the field is still catching up.