US Layoffs Hit Their Highest Level Since the Pandemic — AI Cited as the Cause of 40% of Cuts

According to IBTimes, layoffs in the US have reached their highest level since the pandemic. Centered on the tech industry, a growing number of companies cite AI adoption as a reason for workforce reductions, with an estimated 40% of cuts linked to AI-driven replacement.

What this figure suggests is a shift in which AI is being repositioned from a mere "assistant tool" to a "replacement." That said, blaming layoffs on AI also gives companies a way to signal to investors that they are "improving efficiency," so the actual scale of AI-driven replacement should be viewed with a critical eye.

China Catches Up to Anthropic's Cybersecurity Capabilities — A New Phase in the AI Race

According to the Wall Street Journal, China has caught up with Anthropic's capabilities in the field of cybersecurity. While the US tries to preserve its edge through export controls, this is a symbolic event showing that China is rapidly closing the technology gap.

One of the factors that has long underpinned US AI superiority is the performance gap of frontier models. However, being caught up in the security domain goes beyond raw model performance and could also affect competitiveness in applications and implementation. With the restricted re-release of the Mythos model making waves, the very effectiveness of the regulations is now being called into question.

Claude Code's Safety Classifier Falsely Blocks Legitimate Work — Sharp Spike in Late June

Reports from the Japanese developer community indicate a sharp spike in cases where Claude Code's safety classifier (the cyber check) incorrectly blocks legitimate development work. According to an article on Qiita, around June 27, 2026, a dozen or so issues were filed on GitHub within 24 hours, and none of them involved attacks or violations — they were all ordinary development and research work.

Safety filters are essential for preventing malicious use, but when "false positives" that halt legitimate work pile up, they directly hurt developer productivity. Particularly in contexts like security-related code reviews and penetration testing, tool names and commands tend to trip the filter, and some developers report spending extra time hunting for workarounds.

This problem is not limited to Claude Code — it applies to AI coding tools in general — but as adoption expands, improving accuracy has become an urgent priority.

Running Claude Code "Autonomously" for a Month — Activity and Results Are Two Different Things

Also on Qiita, a developer shared their experience of running Claude Code in a mostly autonomous loop for several weeks. They had a string of lively sessions where "commits piled up, files were touched, and tasks got completed," but looking back at the actual output, the author's conclusion is that "looking busy" and "delivering results" are two very different things.

What makes this article interesting is how it separates "quantitative metrics" — the AI agent's activity logs — from "qualitative results" — features actually merged and issues actually resolved. The lesson that autonomous agents easily create the illusion of "working" but should be measured by their results is practical insight for any team rolling out AI coding tools.

Setting Project-Specific Rules in Antigravity IDE — Putting the Gemini Pro Integration into Practice

A post on Qiita walks through how to configure project-specific rules in Google's Antigravity IDE. Antigravity IDE is a product that was split off as its own development environment (editor) in the May version update and is available with the Gemini Pro plan.

Much like CLAUDE.md in Claude Code, it provides a mechanism for defining coding conventions and architectural assumptions as rules on a per-project basis. As AI-assisted IDEs become one of several options to choose from, "how to convey project context" in each tool is emerging as a shared challenge.

Bash4LLM+ — A Zero-Dependency LLM API Wrapper

On Hacker News, an LLM API wrapper called "Bash4LLM+" written in pure Bash (4+) has been released. It has no dependencies beyond curl and jq, and comes with streaming API support, NDJSON-based session history management, and an interactive REPL mode. It is drawing attention as an approach for using LLM APIs with a minimal footprint — handy on small VPS instances where you want to avoid Python virtual environments or NPM packages.

---