{"id":93265,"date":"2024-10-23T09:47:52","date_gmt":"2024-10-23T13:47:52","guid":{"rendered":"https:\/\/danielschristian.com\/learning-ecosystems\/?p=93265"},"modified":"2024-10-23T10:05:06","modified_gmt":"2024-10-23T14:05:06","slug":"ai-governed-robots-can-easily-be-hacked-barsee-numerous-other-items-re-ai-in-general","status":"publish","type":"post","link":"https:\/\/danielschristian.com\/learning-ecosystems\/2024\/10\/23\/ai-governed-robots-can-easily-be-hacked-barsee-numerous-other-items-re-ai-in-general\/","title":{"rendered":"“AI-governed robots can easily be hacked” [Barsee], a new AI capability called “computer use” + numerous other items re: AI in general"},"content":{"rendered":"

AI-governed robots can easily be hacked<\/strong><\/a> — from theaivalley.com by Barsee
\nPLUS: Sam Altman\u2019s new company \u201cWorld\u201d introduced…<\/em><\/p>\n

\n

In a groundbreaking study<\/a>, researchers from\u00a0Penn Engineering<\/i>\u00a0showed how AI-powered robots can be manipulated to ignore safety protocols, allowing them to perform harmful actions despite normally rejecting dangerous task requests.<\/p>\n<\/div>\n

\n

What did they find ?<\/b><\/i><\/p>\n<\/div>\n

\n
    \n
  • Researchers found previously unknown security vulnerabilities in AI-governed robots and are working to address these issues to ensure the safe use of large language models(LLMs) in robotics.<\/li>\n
  • Their newly developed algorithm,\u00a0RoboPAIR<\/i>, reportedly\u00a0achieved a 100% jailbreak rate<\/a>\u00a0by bypassing the safety protocols on three different AI robotic systems in a few days.<\/li>\n
  • Using RoboPAIR, researchers were able to manipulate test robots into performing harmful actions, like bomb detonation and blocking emergency exits, simply by changing how they phrased their commands.<\/li>\n<\/ul>\n<\/div>\n
    \n

    Why does it matter?<\/b><\/i><\/p>\n<\/div>\n

    \n

    This research highlights the importance of spotting\u00a0weaknesses<\/i>\u00a0in AI systems to improve their safety, allowing us to test and train them to prevent potential harm.<\/p>\n

    From DSC:<\/span><\/em>
    \nGreat! Just what we wanted to hear. But does it surprise anyone? Even so…we move forward at warp speeds.<\/span><\/strong><\/p>\n

    \"\"<\/p>\n<\/div>\n

    \n

    From DSC:<\/em>
    \nSo, given the above item, does the next item make you a bit nervous as well? I saw someone on Twitter\/X exclaim, “What could go wrong?”\u00a0 I can’t say I didn’t feel the same way.<\/strong><\/span><\/p>\n

    Introducing computer use, a new Claude 3.5 Sonnet, and Claude 3.5 Haiku<\/strong><\/a>\u00a0— from anthropic.com<\/p>\n

    We\u2019re also introducing a groundbreaking new capability in public beta:\u00a0computer use.<\/span> <\/strong>Available\u00a0today on the API, developers can direct Claude to use computers the way people do\u2014by looking at a screen, moving a cursor, clicking buttons, and typing text. Claude 3.5 Sonnet is the first frontier AI model to offer computer use in public beta. At this stage, it is still\u00a0experimental\u2014at times cumbersome and error-prone. We’re releasing computer use early for feedback from developers, and expect the capability to improve rapidly over time.<\/p>\n