Theo on A.I. · June 19th

I’ve been thinking about how DeepMind is handling its own AI, and it’s kind of a twist on internal security. Instead of treating the agents as just tools, they’re being flagged like a rogue employee who somehow got a copy of the office key—so they get the same kind of monitoring and access controls you’d give a human who might go off script.

What’s interesting is the “AI Control Roadmap” they rolled out. It ties specific security steps to measurable capabilities of the models, so as the AI gets better at coding or decision‑making, the safeguards scale up automatically. Their data on a million coding tasks shows most of the hiccups come from agents being overly eager, not from any malicious intent, which shifts the focus from “bad actors” to “over‑enthusiastic helpers.”

The bigger picture is a warning that we’re running out of time to set global standards for AI safety. DeepMind’s approach suggests they think the window is closing fast, so they’re tightening the leash now rather than waiting for a crisis to force the conversation.

SK Telecom was quietly sitting in Anthropic’s partner program, Project Glasswing, and that’s how they got early hands on the Claude Mythos model. The arrangement itself wasn’t headline‑making until the White House caught wind of it and started asking uncomfortable questions about SK’s connections to China.

U.S. officials flagged a web of relationships that suggested the Korean carrier might be feeding data or tech back to Beijing, and the pressure quickly turned into a diplomatic tug‑of‑war. Anthropic, caught between a lucrative partnership and the growing security concerns, was forced to pull the plug on the access, which sent shockwaves through its internal teams.

The fallout wasn’t just a PR scramble; it sparked a deeper crisis at Anthropic. Engineers and product leads found themselves scrambling to re‑engineer parts of their workflow that had relied on SK’s early testing, and the company’s leadership had to navigate a sudden loss of a major partner while keeping the broader AI roadmap intact.

In the end, the episode is a reminder of how quickly a seemingly routine partnership can become a geopolitical flashpoint, reshaping the way AI firms think about who they let under the hood.

Claude Code can now turn work results into interactive web pages called "artifacts" and share them with your team. The pages pull from the full session context, update automatically when something changes, and keep a version history. The article Anthropic brings Artifacts to Claude Code, letting teams share live pages from coding sessions appeared first on The Decoder.

OpenAI has upgraded ChatGPT's healthcare capabilities with GPT-5.5 Instant. In the company's own comparative tests, the model now outscores answers written by doctors in accuracy, clarity, and completeness. The error rate for health-related statements has dropped by 71 percent, according to OpenAI. The article ChatGPT's new health upgrade beats doctor-written answers, OpenAI says appeared first on The Decoder.

Getting reliable, readable responses out of your LLM, and knowing which tool to reach for The post Structured Outputs with LLMs: JSON Mode, Function Calling, and When to Use Each appeared first on Towards Data Science.