LLM agent limitations across vulnerability discovery, privacy preservation, scientific review, and browser use; plus on-device cohort inference, SLUB allocator exploitation, temporal HDR corruption, and adversarial DDoS detection.
New findings on learning-based detector poisoning, capability-based access control, and hardware accelerators for post-quantum signatures.
Today’s preprint highlights include unprivileged IMU side-channels on Apple Silicon, TEE-backed digital rights architectures, hardware-rooted Linux anti-cheat, provenance-aware transformers against indirect prompt injection, and cryptographic implementation benchmarking for LLMs.
LLM agents are central today, revealing new architectural risks like capability laundering and the stochastic deputy problem, while hardware vulnerabilities expand into GPU memory architectures.