Daily Research Brief 2026-08-02

📊 Token usage: estimated from retrieval and writing scale.

Covers the latest AI research, open source and industry moves, updated daily.

Editor’s Note

Two signals worth watching today. First, the safety boundary of multimodal agents is spilling from ’text guardrails’ into the ‘perception channel’: on arXiv, ‘Safeguards Based on Copyable Context’ proves a formal trilemma that if evidence can be copied, upload-side safety cannot hold; concurrent audio prompt injections against multimodal agents are now demonstrated.

1. Latest arXiv Papers

1. Safeguards Based on Copyable Context Cannot Provide Reliable Safety for LLMs — https://arxiv.org/abs/2607.27951

2. Piggybacking on Perception: Stealthy Concurrent Audio Prompt Injections against Multimodal LLM Agents — https://arxiv.org/abs/2607.28165

3. Do Context Files Help Coding Agents? A Two-Agent Ablation Study on Real Repositories — https://arxiv.org/abs/2607.27250

4. Beyond KV Reconstruction: Functional Reconstruction for MLA Draft Models — https://arxiv.org/abs/2607.27269

5. GuideSkill: Evolving Executable LLM Agent Skills — https://arxiv.org/abs/2607.26160

6. LabEvolver: Training-Free Experience Evolution for Safe and Grounded Wet-Lab Agents — https://arxiv.org/abs/2607.27690

7. Cross-Embodiment Transfer via Behavior-Aligned Representations — https://arxiv.org/abs/2607.27549

8. Failure Detection for Surgical Robot Imitation Policies via Flow-Matching World Modeling — https://arxiv.org/abs/2607.27511