AI Reading List

Daily curated list of news, articles, and other interesting posts related to AI.

2025-11-20

• Adversarial Poetry as a Universal Single-Turn Jailbreak Mechanism in LLMs (arxiv.org)

  Interesting findings on how converting harmful prompts into poetry dramatically increases the success rate of jailbreaking large language models across many providers. The study highlights a major vulnerability in current AI safety systems, suggesting that models can be reliably tricked by merely changing the style of prompts.

• The AI bubble is bigger than you think (prospect.org)

  Interesting take on how much of the hype around AI is fueled by risky and often opaque financial practices, not just technological advances. Highlights concerns that the mix of Silicon Valley and Wall Street money could be setting the stage for a larger financial bubble.

• Gmail can read your emails and attachments to train its AI, unless you opt out (www.malwarebytes.com)

  Google has started opting Gmail users in by default to allow email and attachment data to train its AI models for features like Smart Compose, raising privacy concerns for those who handle sensitive information. People who want to prevent their emails from being used for AI training need to manually turn off smart features in two different settings screens.