AI Safety and Cybersecurity (Links) – May 12, 2026

Written by

in

, ,
  • Simon Willison: A quote from New York Times Editors’ Note (May 10, 2026)
    This article was updated after The Times learned that a remark attributed to Pierre Poilievre, the Conservative leader, was in fact an A.I.-generated summary of his views about Canadian politics that A.I. rendered as a quotation.”
  • Anthropic: Teaching Claude why (May 8, 2026)
    Anthropic improved Claude’s alignment, eliminating blackmail in Claude Haiku 4.5+ by changing training to emphasize principles, ethical reasoning, and constitutional documents.
  • Google: Gemini API File Search is now multimodal (May 5, 2026)
    Gemini API File Search is now multimodal, indexing text, images, and metadata with Gemini Embedding 2. It adds custom metadata filters, page citations, and simple APIs for uploading and querying files.
  • Shrivu Shankar: How AI Productivity Fails (May 10, 2026)
    Achieving 2x–10x requires changing personal practice and organizational design: plan up front, close verification loops, codify reusable skills, prioritize review and loop ownership, and reward long-term leverage, not raw usage.
  • Tyler Cowen: Will AI kill the research paper? (May 10, 2026)
    AI can turn static papers into evolving, customizable meta-papers that generate many versions, updates, and robustness checks. Research will shift to building and maintaining these boxes.
  • Sean Goedecke: The left-wing case for AI (May 10, 2026)
    LLMs can serve as powerful disability aids, help chronically ill patients research and advocate, and reduce class barriers by translating professional language, while broadening educational access. 
  • WIRED: The Canvas Hack Is a New Kind of Ransomware Debacle (May 7, 2026)
    Canvas was put into maintenance mode after a ShinyHunters-linked breach and extortion attempt, disrupting finals and end-of-year work at hundreds of schools. Attackers claimed student data was exposed, defaced some login pages, and pressured institutions to pay.
  • Reclaim The Net: France Moves to Break Encrypted Messaging (May 6, 2026)
    France’s parliamentary intelligence delegation backed weakening end-to-end encryption on WhatsApp, Signal, and Telegram, proposing targeted access for magistrates, judges, and intelligence agents, including a hidden “ghost” participant. Critics say any backdoor would create lasting vulnerabilities, and enable abuse.

Comments

Leave a Reply

Your email address will not be published. Required fields are marked *