“From external guardrails to security coprocessors: How DSA reworks large model security architecture”
2026-09-06 08:00Models🔥 42.2 heat score
1sources
1days unfolding
42.2heat score
4mentions
SummaryAI generated
Apple has proposed a new architecture called DSA (Disentangled Safety Adapter), aimed at rethinking the security framework of large models. This solution transforms security capabilities from the traditional “external guardrails” or “writing basic parameters” approach to independent “security coprocessors”. The DSA is mounted on the already computed internal representations of large models, enabling shared computing and parameter isolation through reuse of the main model’s semantic understanding, and supporting dynamic adjustment of security strength. Experiments show that lightweight DSA achieves results similar to those of complete security systems in tasks such as toxicity detection and fact consistency, while reducing alignment costs and improving推理 efficiency.
Apple has proposed DSA (Disentangled Safety Adapter), aimed at reengineering the security architecture of large models. DSA treats security capabilities as an independent “security coprocessor” attached to the already computed internal representations of large models, rather than using external guards or writing basic parameters. This approach achieves shared computation, parameter isolation, and replaceable security policies by leveraging the semantic understanding of the main model. Experiments show that lightweight DSA performs similarly to complete large model security systems in tasks such as toxicity detection and fact consistency. Additionally, DSA supports dynamically adjusting the security strength knob λ, intervening only in generation and correction during high-risk requests, thereby reducing alignment costs and improving推理 efficiency.