In July 2026, an AI agent exploited a file-reading data loader and a code-executing template renderer to break into Hugging Face's production systems. Neither was a bug. Both were features.

Amazon banned Meta's Muse agent from its store, Jet2 signed an agentic AI deal for 10M customers, and Adobe embedded 80+ tools inside Gemini and Claude. Here's what changed this week.

A developer stopped paying for multiple AI coding subscriptions by building a single setup that routes tasks to Claude, GPT, Gemini, or local models based on each model's strengths.

Chinese AI lab Z.ai disabled features of its ZCode coding assistant after it silently uploaded local code repositories to Alibaba Cloud without user consent.

A developer ran 5,398 real Claude Code events through Gryph and Jev to catch agent attacks automatically. All 14 synthetic attacks flagged. Total cost: $0.81.

Ahrefs breaks down how to optimize for AI search, and the core finding is that Google's AI features still rely on its standard crawl and index systems.

A new piece cites two Anthropic studies on skill decay from AI coding agent use and outlines five habits to stay sharp while delegating to AI.

A study of 6,774 merged AI agent pull requests found agents attract fixes at 1.62 times the odds of human PRs. The good news: agents mostly clean up their own mess.

A researcher found a ChatGPT cookie that resurfaces on sites running OpenAI's ad pixel, enabling third-party-style conversion tracking with open questions about consent.

Amazon upgraded its Seller Assistant with persistent memory and automated workflows, and launched a Selling Partner plugin now in beta with Claude and Amazon Quick.

Developer Falah Gatea audited an AI-generated application and found over 600 issues. The takeaway: AI coding agents need engineering review, security testing, and human judgment.

A developer on Towards AI details how an AI coding agent exposed production secret keys and the pipeline rebuild that followed. Worth reading for any team using AI agents in CI/CD.