AI & Cybersecurity: Dan Boneh Interviews Nicolas Carlini
- Source
- youtube.com
- Author
- Stanford Online
- Date
Context
Hiding a 's weights (its learned numerical settings) does not reliably protect it from adversarial examples, because inputs crafted against one model often work on another, Nicholas Carlini tells Stanford's Dan Boneh in this interview. An adversarial example is an input altered to make a model misbehave, such as a picture changed so a classifier labels a cat as guacamole. Common attack methods need the weights to compute the alteration, so secrecy sounds protective. Carlini says an attacker can build the example on a different model, trained on different data, and replay it on the target. It works more often when architectures and data are similar, and querying the target to adapt improves the odds.
Carlini is explicit about the limit: transfer attacks do not always succeed, so secrecy raises an attacker's cost, as spam defenses do, without delivering robustness. Carlini, whom the interview's introduction places at Anthropic working on security, adds that such attacks matter more now that language models make real decisions, such as screening uploads.
- Model — A computational system that uses learned patterns to produce an output from information supplied to it.
- Training — The process that adjusts a model using examples and a measure of how well it performs.
- Large language model — A model trained on large amounts of data to process and generate language.
videoStanford Webinar - Building Trustworthy AI: Navigating Security in the Agentic Era
videoStanford CS547 HCI Seminar | Autumn 2025 | HCI Systems in the Age of AI Code Generation
videoStanford CS25: Transformers United V6 I Advancing Science and Medicine with Collaborative AI Agents
videoStanford CS153 Frontier Systems | Andreas Blattmann from Black Forest Labs on Visual Intelligence
Checking sign-in…
Loading comments…


