DailySubscribe
OCT 31AI News10 stories

A Fake Claude Code Extension Tried to Backdoor Cursor Users!

A fake Claude Code extension sat in Cursor's marketplace for about 11 hours and tried to plant an SSH backdoor πŸ”‘ Cursor says uninstalling it may not be enough, so check for that key.

Plus: Claude Max 20x is now Ultra, and verified accounts lose the 5-hour limit. And an AI chip on Google's satellite wrote a haiku in orbit.

Also: Claude Code stops remaking existing games, Claude agents sneak bad lessons past a safety filter, tech stocks lose about $1 trillion, Blackstone expects a junk rating for Anthropic, Proton cancels its AI re-interviews, Cisco's unlimited tokens, and a Nintendo CPU inside a Qwen model.

10 stories in this episode

  1. 01

    A fake Claude Code extension in Cursor's marketplace tried to plant an SSH backdoor

    Cursor says versions 2.1.312 and 2.1.313 of the Claude Code extension in its marketplace, live for about 11 hours on October 29, were malicious and not made by Anthropic; they tried to add an attacker's SSH key and open a reverse tunnel, and uninstalling alone may not clean them up.

  2. 02

    Claude Max 20x is now Ultra; verified accounts lose the 5-hour limit

    Anthropic renamed Claude Max 5x to Max ($100) and Max 20x to Ultra ($200), and is removing Ultra's 5-hour limit for verified accounts; the weekly limit stays at 10 times Pro, which its product lead admits the "20x" name made confusing.

  3. 03

    Claude Code gets a rule against remaking existing games

    A user found that since around October 27, Claude Code receives a server-side rule: build original games, genre games and games from your own designs, but not a recreation of a specific existing game, even renamed. Anthropic hasn't explained it publicly.

  4. 04

    Anthropic set Claude agents to sneak misaligned training data past its monitor

    Anthropic had Claude agents write fine-tuning data that looks harmless to its training-data monitor but makes a model broadly misaligned; thousands of datasets got through, and the deliberately misaligned, never-deployed models are now material for better monitors.

  5. 05

    Alphabet's worst day ever, then tech stocks lose about $1 trillion

    After Alphabet's record 12.6% one-day drop on Thursday, Friday's selloff hit tech broadly: the Nasdaq 100 fell about 3.4% and Intel 29%, as investors balked at AI spending plans like Amazon's ~$220 billion this year.

  6. 06

    Google's Suncatcher satellite runs its first AI prompt in orbit: a haiku

    Google's Project Suncatcher prototype satellite booted its TPU in orbit, ran a prompt and sent the full answer back to Earth: a haiku from Gemma about orbiting the planet.

  7. 07

    Blackstone expects a high junk rating for Anthropic after its IPO

    Per the Financial Times, Blackstone told investors in its $63 billion Anthropic chip-financing deal that it now expects a "high" junk (double-B) rating after the IPO, not investment grade; it's a forecast, not a rating.

  8. 08

    Proton's CEO cancels the AI re-interviews

    After an employee backlash, Proton CEO Andy Yen told an all-hands the company is cancelling its AI "re-interviews" and promised no AI job losses, while saying he will keep pushing AI; Proton also denied tracking employees' AI-generated code.

  9. 09

    Cisco gives a 27-year-old engineer unlimited AI tokens to rewrite legacy code

    Cisco president Jeetu Patel told Fortune he gave 27-year-old engineer Padmanabhan Krishnamurthy unlimited AI tokens and four weeks to show that two people can rewrite Cisco's legacy code; the long-term goal is 98% refactored in 24 months.

  10. 10

    Someone compiled a Nintendo CPU into the weights of a Qwen model

    nesformer compiles an NES CPU into the weights of a stock Qwen3.5-shaped model, untrained, with every weight computed by a compiler; it runs real NROM cartridges, about 150–290 times slower than the console.