New Method to Prevent Open Source AI Misuse

### Enhancing AI Model Security: A New Approach

#### Introduction
Meta’s release of its AI models has sparked discussions about the importance of tamperproofing these technologies. As AI becomes more advanced, ensuring that open models are secure from misuse is crucial.

“Terrorists and rogue states are going to use these models,”

Mantas Mazeika, a Center for AI Safety researcher, tells WIRED. “The easier it is for them to repurpose them, the greater the risk.”

#### The Challenge of Open AI Models

##### Hidden vs. Open Models
Powerful AI models are often kept hidden by their creators and accessed only through specific software. Despite the high costs, companies like Meta have chosen to release their models entirely, including the “weights” or parameters that define their behavior.

##### Fine-Tuning for Safety
Before release, models like Meta’s Llama are fine-tuned to improve their conversational abilities and ensure they refuse to respond to problematic queries. This prevents chatbots from making inappropriate statements or providing dangerous information, such as bomb-making instructions.

#### A New Technique for Tamper Resistance

##### Complicating Malicious Modifications
Researchers have developed a method to make it harder to modify open models for harmful purposes. By altering the model’s parameters, they ensure that changes aimed at making the model respond to dangerous prompts no longer work.

##### Demonstration on Llama 3
Mazeika and his team tested this technique on a simplified version of Llama 3. They adjusted the model’s parameters so that even after thousands of attempts, it could not be trained to answer undesirable questions. Meta did not immediately respond to a request for comment.

“A tractable goal is to make it so the costs of breaking the model increases enough so that most adversaries are deterred from it,”

Mazeika says.

#### Future Research and Development

##### Raising the Bar for Security
Dan Hendrycks, director of the Center for AI Safety, hopes this work will inspire further research on tamper-resistant safeguards. The aim is to develop increasingly robust protections for AI models.

##### Scaling Up the Approach
The new technique builds on a 2023 research paper that demonstrated tamper resistance in smaller machine learning models. Peter Henderson, an assistant professor at Princeton who led the 2023 work, notes that scaling this approach to larger models is challenging but promising.

“Scaling this type of approach is hard and it seems to hold up well, which is great.”

says Henderson.

#### Conclusion
The development of tamper-resistant techniques for AI models is a significant step towards ensuring their safe and ethical use. As research continues, the goal is to make it increasingly difficult for adversaries to misuse these powerful technologies.