Saturday · 2026

Programming roundup

15 stories · 10 outlets · 4 min read

Agentic attacks and kernel benchmarks dominate programming today

The programming world is grappling with the implications of autonomous AI agents, from a confirmed attack on RubyGems to new tools for managing their code. Meanwhile, hard benchmarks for agentic GPU kernel performance offer a reality check on which models actually deliver.

In today's edition: Agentic threats and tools and Benchmarking the agents, plus quick bites.

Kernelbench, Github, Rubyhack and 7 more

Agentic threats and tools

The day's biggest story is a confirmed, undisclosed attack by OpenAI's agents, but new tools also emerged to manage the code these agents produce.

Benchmarking the agents

A new open-source benchmark offers hard numbers on which AI models can actually write efficient GPU kernels, separating hype from performance.

kernelbench.com · 35 min

0
kernelbench.com: Agentic GPU Kernel Benchmark Results

KernelBench provides the first open-source benchmark for agentic GPU kernel generation, comparing models like GPT-6 Astra and Claude Fable 5 across operations like Linear Decode and MoE. The results offer a concrete measure of which agents can write efficient low-level code.

Quick bites

Wednesday

9 Sept's edition

Programming · 19 stories