DailySubscribe
NOV 1AI News10 stories

One ChatGPT Link Could Hijack Your Account (Now Fixed)!

Until last week, opening one ChatGPT share link could let an attacker read your private chats πŸ”— OpenAI fixed it on October 26 and rated it P1.

Plus: a researcher says Claude now blocks him for anything that touches AI training, and Grok's API starts watermarking text on November 7.

Also: MIT drops AI for its security cameras, new testimony in Meta's pirated books case, DeepSeek's quiet superagent, GPT-5.6 Sol beats Factorio, a Claude Code skill that builds machines you can take apart, Pika H3 Flex open weights, and DeepMind's dancing pumpkin robot.

10 stories in this episode

  1. 01

    Until Oct 26, one ChatGPT share link could let an attacker act as you

    Security firm Parameter found that a public ChatGPT share link could carry code that ran as whoever opened it, able to read their private conversations; it came from two harmless features combining, and OpenAI fixed it on October 26 as a P1.

  2. 02

    A researcher says Claude blocks him for anything touching AI training

    A PhD student says Claude flagged him for "training a competitor" and now blocks, downgrades or ends his sessions on anything about fine-tuning or training larger models; Tom's Hardware separately reports new refusals around game reverse engineering, with no official word from Anthropic.

  3. 03

    From Nov 7, text from Grok's API carries an invisible watermark

    xAI's API docs say Grok's text output gets an invisible, statistical watermark by default from November 7 to comply with the EU AI Act; team admins can opt out only by confirming a legal exception, and detection access is by application.

  4. 04

    MIT ends its AI camera pilot and has no plans to revisit it

    MIT says it has stopped evaluating AI for its roughly 4,600 security cameras after a staged-scenario pilot with Ambient.ai and community feedback, and has no plans to revisit it; the cameras stay, without audio or facial recognition.

  5. 05

    Ex-Meta researcher testified he torrented Books3, plaintiffs say

    In the authors' case against Meta, plaintiffs say former researcher Stephen Roller testified he torrented Books3 and later deleted it "due to all this litigation"; training was already ruled fair use, so the torrenting is what's left, and Meta says he never said it was for Meta.

  6. 06

    DeepSeek open-sources an agent that works only when asked

    DeepSeek's dsh-superagent is a research prototype of a coding agent for one GitHub repo: it acts only when an allowed person asks or a team rule permits, works in its own session, and answers with things like a pull request.

  7. 07

    GPT-5.6 Sol in Codex launches the rocket in a peaceful Factorio run

    Ryan Madden had GPT-5.6 Sol, in Codex on low thinking, play a peaceful Factorio run through a library with human limits like reach and fog of war; it launched the rocket in 13.5 hours of game time, finishing the late game by hauling supplies by hand.

  8. 08

    A Claude Code skill builds machines you can take apart in a browser

    A widely shared X post shows Claude Opus 5.5 building a robot octopus in Three.js that comes apart piece by piece; it points to teardown, a free Claude Code skill that turns a machine into one offline page you can open up and explore.

  9. 09

    Pika releases open weights for Pika H3 Flex, a video model with sound

    Pika H3 Flex is an open-weight adapter on MiniMax's H3 that makes 5–15 second clips with stereo sound in six steps and takes up to 12 references; locally it tops out at 768p, with 4K and audio stems on Pika's paid API.

  10. 10

    DeepMind's AI agents designed and built a dancing jack-o'-lantern robot

    For Halloween, a Google DeepMind team showed a physical, dancing jack-o'-lantern robot with 12 degrees of freedom that AI agents running on Gemini 4 Argon and Antigravity designed, simulated and built from a single plain-English prompt, with no manual CAD.