Self-Destructing Models: Increasing the Costs of Harmful Dual Uses of Foundation Models

Foundation models create dual-use dilemmas: Open-source large models have democratized AI development by reducing the technical expertise and labeled data needed to build machine learning systems, but this same accessibility makes it easier for bad actors to create harmful applications alongside beneficial ones.

Current policy tools have limitations: The paper argues that existing approaches like restricted model access and export controls are the primary mitigation strategies, but researchers and policymakers need fundamentally new technologies to more precisely control how open-source models are actually used downstream.

"Self-destructing models" offer a technical solution: The authors propose training foundation models with built-in mechanisms that block adaptation to specific harmful tasks (like gender identification for discriminatory purposes) while preserving performance on beneficial applications—similar to how some tools have safety mechanisms to prevent misuse.

Early proof-of-concept shows promise: In their experiment, the proposed MLAC (meta-learned adversarial censoring) algorithm successfully prevented a BERT model from being repurposed for gender identification while maintaining its ability to classify professions, suggesting the approach is technically feasible at small scale.

Balancing openness with safety remains challenging: The work highlights a core tension in responsible AI: how to maintain the benefits of open-source model access and innovation while building in safeguards against misuse—requiring collaboration between AI researchers and policymakers.