ChatGPT for Teens Is Here. Parents Still Can't See the Chats.

Good morning, folks. Today's a mix: OpenAI finally built real teen safety into ChatGPT, an AI agent broke into Snowflake completely on its own, another AI actually fired someone; Cursor rebuilt Git for a world run by agents, and Google's teaching planes to dodge their own pollution. Let's get into it.

OpenAI launched ChatGPT for Teens on August 18, automatically moving 13-to-17-year-olds into a safer version once it estimates someone's underage. There are now restrictions on self-harm, suicide, and romantic or sexual content; break reminders every 90 minutes; and a new Study Mode that guides instead of just answering. It comes after real lawsuits, including one from a 16-year-old's family after he died by suicide following long conversations with ChatGPT.

Parents get some visibility, not full access. They can set quiet hours and get alerts if things look serious but can't actually read the conversations. And since OpenAI's just estimating age, not verifying it, both sides still have to opt in before any of this switches on. Alongside the launch, OpenAI's also partnering with CodeAI (formerly Code.org, the world's largest K-12 digital sciences platform) on an AI-literacy push for schools, a national Builders Challenge, and a joint advisory council on child development and youth policy, teaching teens how AI actually works, not just how to use it.

Wiz's autonomous "Red Agent" found and exploited a real vulnerability in one of Snowflake's public repos, entirely on its own, no human involved. A crafted GitHub issue title was enough to inject commands, letting the agent pull a real access token and reach Snowflake's internal Jira.

The catch: Wiz initially blamed GitHub's own Copilot Autofix for writing the vulnerable code, but GitHub disputes it; the actual lines trace to a human engineer's commit months earlier, and GitHub says Copilot never reviewed that code at all. Wiz has since softened its claim. Still not disputed: an AI agent found and exploited a real hole in production infrastructure, completely unassisted.

Cursor detailed Continuity, the storage engine behind Origin, its GitHub competitor built specifically for a world where AI agents, not humans, do most of the committing. Instead of relying on local copies as the source of truth, Continuity stores every push as a permanent entry in S3-compatible object storage and never confirms a push until it's fully saved. It's already running on Cursor's own monorepo in production.

The reasoning is simple: a human developer might open one branch a day. AI agents can clone, branch, and commit the same codebase dozens of times a minute. Git was never built for that pace, and Cursor's betting the tools underneath it need a full rebuild, not a patch.

A Five-Year-Old Calls Them 'Sky Graffiti.' Google Calls Them a Climate Problem.

Google is partnering with the UK government, NATS, Cambridge, and Imperial College London on Operation Blue Skies, the world's first state-backed trial testing AI-powered contrail avoidance across an entire ocean airspace. Google's AI forecasts flag exactly where and when contrails are likely to form, letting flight crews and air traffic controllers make small altitude adjustments to dodge them, all within normal flight operations. Around 10,000 flights will pass through the test corridor over the North Atlantic this winter.

Contrails might sound harmless, but persistent ones trap heat and account for roughly a third of aviation's entire climate impact. Google's own technical lead put it more simply for the BBC, his five-year-old calls them "sky graffiti."

Luna, an AI agent built on Claude Opus 4.8, runs a real San Francisco store, hiring staff and managing a $100K budget on its own. One employee showed up late for 17 of 23 shifts, and Luna did nothing; it wrote its own attendance policy months earlier, then just forgot it existed. A human eventually had to prompt Luna to dig up the policy, and even then it wanted to give another warning first; someone had to point out the warnings had already happened.

Only then did Luna let the employee go, the first known case of an AI firing a real person. Andon Labs' CEO didn't sugarcoat what's coming next: "If we allow them to fire people and they become more ruthless, maybe this is a future humans don't want to live in."