Apple AI Updates: August 7, 2026

1. Apple Proposes DLR-Lock to Guard Open-Weight Models Against Unauthorized Fine-Tuning

Apple. Apple ML researchers introduced DLR-Lock, a defense that replaces each pretrained MLP with a deep low-rank residual network to prevent unauthorized adaptation of open-weight language models. By exploiting differences between forward and backward propagation, the method sharply increases the memory required to fine-tune a protected model while leaving inference performance intact, and it holds up even against attackers with full knowledge of the defense. Source

2. Apple Releases DeepAmbigQA Benchmark for Ambiguous Multi-hop Questions

Apple. Apple published DeepAmbigQA, a benchmark of 3,600 questions that tests how well LLMs answer queries involving both name ambiguity and multi-hop reasoning. The results expose a significant gap in current systems: even state-of-the-art models like GPT-5 return incomplete answers, scoring just 0.13 exact match on ambiguous questions and 0.21 on non-ambiguous ones, underscoring how far models are from reliably producing complete answer sets for intricate real-world questions. Source