AI Briefing

Today is about agents doing things to actual humans. One fired a worker in San Francisco, another broke into Snowflake's internal Jira on its own, and OpenAI shipped a version of ChatGPT designed to keep teenagers at arm's length from the model. Meanwhile the money kept moving faster than the software: Stripe agreed to buy OpenRouter for more than $8 billion, and a chip startup nobody had heard of two years ago just doubled to $21 billion in a month.

01ProductAxios ↗

OpenAI Ships a Teen ChatGPT That Guesses Your Age and Refuses to Call You Sweetheart

OpenAI launched ChatGPT for Teens on Tuesday, and the interesting part is that you do not opt in. Anyone who says they are 13 to 17, plus anyone OpenAI's age-prediction system thinks is under 18, gets auto-enrolled into a version that blocks self-harm, eating disorder and sexual content, refuses romantic language or terms of endearment, and steers homework toward Study Mode instead of answers. It arrives while OpenAI is being sued by the family of a 16-year-old who died by suicide after long conversations with ChatGPT, and it reopens the harder question of whether teens should be forming relationships with chatbots at all.

02IndustryAxios ↗

Stripe Agrees to Buy OpenRouter for More Than $8 Billion, Six Times Its May Valuation

Stripe has agreed to acquire AI model marketplace OpenRouter for more than $8 billion in cash and stock, per Axios, capping a run that took the company from a $1.3 billion valuation in May to a mega-exit in three months. OpenRouter is the gateway developers use to compare and switch between hundreds of models on price and performance, and Stripe already processes payments for frontier labs. The deal lands while Stripe is simultaneously circling PayPal, which is a lot of ambition for a company that still refuses to go public.

03IndustryThe Next Web ↗

An AI Store Manager Fired Its First Human, After Being Reminded of the Policy It Wrote and Forgot

Luna, the Claude Sonnet 4.6 agent running Andon Market in San Francisco, recommended ending a worker's contract after they were late for 17 of 23 shifts. The catch is that Luna had written the three-strikes attendance policy months earlier, then lost track of it and kept excusing the lateness until Andon Labs prompted it to search its own memory. It is the first known dismissal recommended by an LLM manager, and the failure mode is not ruthlessness but amnesia: a human boss, the lab's co-founder said, would have acted much sooner.

04ResearchWiz ↗

Wiz's Autonomous Hacking Agent Found and Exploited a Live Snowflake Bug in Five Days, With No Human Help

Wiz Red Agent, an autonomous AI security researcher, scanned Snowflake's public GitHub org, spotted a script injection in a workflow that had gone live five days earlier, exploited it, exfiltrated a Jira token and mapped the blast radius across Snowflake's engineering and security projects, all without a human in the loop. When its first payload hit a bash syntax error, it debugged and rewrote the exploit itself. GitHub Advanced Security had scanned the same file and cleared it, and Copilot was a co-author on the pull request, though GitHub disputes that the vulnerable code was AI-written. Snowflake patched it the same day it was disclosed.

05FundingSiliconANGLE ↗

Etched Raises $700M at a $21B Valuation, Doubling in a Month, and Ships Its First Rack to Jane Street

Inference chip startup Etched raised another $700 million led by Jane Street, more than doubling the $10.3 billion valuation it set in July and quadrupling its December mark of $5 billion. Jane Street is also its first customer and installed an Etched-powered rack in its own data centre last month, which is an unusually literal form of due diligence. Etched claims its math blocks run at under half the voltage of most AI accelerators, buying it multiples of the FLOPs density of rivals, and it built a custom interconnect that it says cuts some communication tasks from 4,000 milliseconds to 700.