JSON

QuoteBench is a new benchmarking framework that exposes how the execution environment of LLM agents can corrupt Bash commands through parsing errors. It demonstrates that high 'matched' success scores often hide significant failures occurring at the boundary between model output and shell execution.

This research evaluates whether replacing structured JSON tool calls with executable Python scripts improves LLM agent performance. The findings suggest that programmatic tool calling is a more robust and scalable alternative, particularly for complex, multi-step, and high-parallelism tasks.

A heap buffer overflow in MariaDB's JSON schema validation allows authenticated attackers to escalate privileges and execute arbitrary code. Users should upgrade to the patched versions 11.4.10 or 11.8.6 immediately.