AI Business LensTHE BUSINESS OF AI, FOR PEOPLE WHO TEACH IT OR LEARN FROM IT
Schneier on Security · October 27, 2026 · On the brief until November 10, 2026

Hiding Capabilities in AI Models

CybersecurityComputer Science
THE AI ANGLE
Hiding encrypted capabilities inside model weights

Researchers demonstrated that backdoor attacks can embed encrypted neural networks inside an AI model's weights using digital lockers. These hidden computations execute only after receiving a secret trigger and resist techniques like fine-tuning, weight noising, and gradient searches. The findings show that pre-deployment capability evaluations cannot reliably catch hidden adversarial tasks.

Summary written by AI Business Lens with an AI model from the article at schneier.com. It is not the article, and the publisher has not reviewed it. For publishers.

THE TEACHING ANGLE
Auditors cannot find hidden model functions with standard elicitation tests when those functions sit behind cryptographic locks.

Read the original at schneier.com   Generate teaching or study materials

Instructors get discussion guides, assignments, and mini-cases. Students and readers get a plain summary, points to raise, an exercise, and a self-check. All built from the full article. A free account saves stories and gives you three generations. Stories stay on the brief for 14 days. After that this page keeps the link to the original.

More in Cybersecurity