Module 6 of 6
Frontier model security
Protecting the weights of highly capable models from theft, insiders, and extraction, and tying release decisions to measured capability.
The more capable a model is, the more its weights are worth to someone else and the more harm a stolen copy can do without the safeguards that surround the official deployment. This module treats the weights as the asset: where copies live, who and what can read them, how they can leave, how much an API gives away, how capability evaluations decide which protections are required, and how hardware isolation keeps weights and prompts away from the people who run the machines.
By the end of it you should be able to place a system on the RAND security levels, inventory and narrow the paths to a model’s weights, design a two-person access broker, size an egress budget and an extraction budget, wire a capability threshold into a release gate, and explain what an attestation report proves before a key is released.
Lessons
- 6.aSecuring model weightsWhat does it take to keep a frontier model's weights from being stolen?15 min
- 6.bResearch and training infrastructureHow do you protect weights in the environments where researchers actually work?12 min
- 6.cInsider risk and model accessHow do you let people work with a model without letting any one of them walk away with it?13 min
- 6.dExtraction and distillationHow much of a model can be copied through its own API, and what raises the cost?14 min
- 6.eCapability evaluations and release gatesHow do capability evaluations decide which safeguards a model needs, and how is that enforced?13 min
- 6.fConfidential inferenceHow can a model run on hardware whose operator cannot read the weights or the prompts?14 min