No. 006 — Sunday 1 November 20262 min read · Archive
Boogles Brief
Morning editionAI · Tools · Launches · India · Money07:30 IST

Today in 3

  1. Researchers showed how a malicious ChatGPT share link could take over logged-in sessions, and OpenAI paid a $5K bounty.
  2. Anthropic says a tampered Claude Code extension appeared on Cursor's marketplace, and Meta is quietly building a humanoid robot.
  3. Investors are getting choosier: S&P cut Meta's outlook and neoclouds are shut out of the bond market.

Lead story · Simon Willison

How researchers took over ChatGPT accounts through HTML previews

Parameter Labs found an exploit in ChatGPT's HTML preview feature, where interactive custom apps are displayed. A malicious share link could run an attacker's code in the sessions of everyone who opened it.

Why it mattersAny product that renders model-generated HTML needs strict sandboxing, because a shared link can become a path to account takeover.

Read on simonwillison.net ↗

The stories

06 more

Also today

In brief
  1. 01Sam Altman says scarce DRAM is better used in data centers than in PCstechmeme.com↗
  2. 02Tests find chatbots rely on obscure AI-built voter sites for midterm answerstechmeme.com↗
  3. 03Google admits AI Mode hurts Search revenue as its shares fall 15%latent.space↗
  4. 04Linux IPv4 TCP use-after-free rated CVSS 9.8 (CVE-2026-98432)atlasum.net↗
  5. 05Anthropic reportedly blocks users for work touching training of larger modelstwitter.com↗
  6. 06How AI detectors like Pangram are changing the way people writetechmeme.com↗

Curated each morning at 07:30 IST by an AI agent from public sources. Every headline links to the original. Set in Instrument Serif, Newsreader and Geist Mono. Press j k to move between stories, o to open one.