It's a Bird, It's a Plane, It's A.I.

The talk opens with a tour of AI history, using books and films to separate general artificial intelligence from narrower systems built for specific jobs. The speaker argues that security work sits somewhere between art and science, and that practitioners should work alongside these tools rather than dismiss them. Early material explains how large language models predict the next word from training data, with human feedback shaping the results.

The middle section turns to practical concerns and tooling. Bias and hallucinations are presented as limits to plan around rather than reasons to ignore the technology. Open options are highlighted, including a local model runner, a public model hub, and a framework for linking models to external APIs, as ways to keep data off hosted services. Live demos follow: a plain chatbot that reveals its provider, a persona-driven bot, a bot grounded with event details through retrieval, and one that calls a file-hash reputation lookup through a function interface. A final demo has a model review C code for a buffer overflow and suggest a fix.

The closing point is that every integration is dual-use. Exposed functions and APIs can leak data or be abused if no access controls sit between the model and the backend, and attackers can run unrestricted local models to produce malicious code. For practitioners, the takeaway is to treat LLM-connected features as part of the application attack surface, test them like any other input path, and build hands-on familiarity with these tools now.