Why an AI Harness May Matter More Than the Latest Model
Technologist Bruce Schneier on Local Models, AI Harnesses and Corporate Power
Artificial intelligence models have grown highly capable of finding, exploiting and patching software vulnerabilities. But government controls and restricted access won’t contain those capabilities, said security technologist Bruce Schneier.
Schneier said security leaders should look beyond model brands and focus on the software harness surrounding each system. Harnesses manage prompts, outputs, context, guardrails and coordination among multiple models. They can also route tasks between expensive frontier models and cheaper local systems. This architecture may matter more than selecting one leading model.
“The way to think of it is the harness is all the code between what you type, what the model gets, what the model produces and what you see, and that’s a lot of non-AI code,” said Schneier, an adjunct lecturer at Harvard Kennedy School at Harvard University and the Munk School of Global Affairs and Public Policy at the University of Toronto. “It does seem like the models are largely interchangeable.”
He also warned that concentrated corporate power, weak model integrity and poorly designed incentives may create greater risks than access to any single model. As smaller systems approach frontier-model performance, organizations may deploy more AI locally for greater speed, control, privacy and cost efficiency.
In this video interview with ISMG, Schneier also discussed:
- Why export bans can’t control widely available AI capabilities;
- How harnesses, guardrails and multi-model coordination shape performance;
- Why CISOs should examine vendor claims, corporate incentives and model integrity.
Schneier is a security technologist, author and public-interest cybersecurity researcher. He teaches at Harvard Kennedy School at Harvard University and the Munk School of Global Affairs and Public Policy at the University of Toronto, focusing on security, technology policy, AI and the societal effects of digital systems.
Categories: Recorded Interviews, Video