Your AI Watched the Demo and Started Clicking

Podcast: AI Breakdown · Varun Sirohi

Opublikowano: 2026-09-24

𝐀 𝐥𝐚𝐮𝐧𝐜𝐡 𝐯𝐢𝐝𝐞𝐨 𝐝𝐢𝐝 𝐚 𝐡𝐮𝐧𝐝𝐫𝐞𝐝 𝐚𝐧𝐝 𝐭𝐡𝐢𝐫𝐭𝐲 𝐭𝐰𝐨 𝐦𝐢𝐥𝐥𝐢𝐨𝐧 𝐯𝐢𝐞𝐰𝐬 𝐭𝐡𝐢𝐬 𝐰𝐞𝐞𝐤: 𝐚 𝐩𝐞𝐫𝐬𝐨𝐧 𝐭𝐚𝐥𝐤𝐢𝐧𝐠, 𝐡𝐚𝐧𝐝𝐬 𝐨𝐟𝐟 𝐭𝐡𝐞 𝐤𝐞𝐲𝐛𝐨𝐚𝐫𝐝, 𝐰𝐡𝐢𝐥𝐞 𝐭𝐡𝐞 𝐀𝐈 𝐥𝐢𝐬𝐭𝐞𝐝 𝐢𝐭𝐞𝐦𝐬 𝐟𝐨𝐫 𝐬𝐚𝐥𝐞, 𝐫𝐞𝐯𝐢𝐞𝐰𝐞𝐝 𝐚 𝐜𝐨𝐧𝐭𝐫𝐚𝐜𝐭, 𝐨𝐫𝐝𝐞𝐫𝐞𝐝 𝐟𝐨𝐨𝐝 𝐚𝐧𝐝 𝐛𝐨𝐨𝐤𝐞𝐝 𝐚 𝐭𝐞𝐧𝐧𝐢𝐬 𝐜𝐨𝐮𝐫𝐭. 𝐓𝐰𝐨 𝐬𝐩𝐞𝐧𝐝 𝐦𝐨𝐧𝐞𝐲. 𝐎𝐧𝐞 𝐢𝐬 𝐚 𝐥𝐞𝐠𝐚𝐥 𝐫𝐞𝐯𝐢𝐞𝐰. 𝐍𝐨𝐛𝐨𝐝𝐲 𝐩𝐫𝐞𝐬𝐬𝐞𝐝 𝐜𝐨𝐧𝐟𝐢𝐫𝐦. 😳 The chat window just grew a pair of hands. And the failures that brings make no noise. This week on 𝐓𝐡𝐞 𝐀𝐈 𝐁𝐫𝐞𝐚𝐤𝐝𝐨𝐰𝐧, Quentin and Stark sharpen the theme: 𝐖𝐡𝐚𝐭 𝐡𝐚𝐩𝐩𝐞𝐧𝐬 𝐰𝐡𝐞𝐧 𝐀𝐈 𝐬𝐭𝐨𝐩𝐬 𝐚𝐬𝐤𝐢𝐧𝐠 𝐩𝐞𝐫𝐦𝐢𝐬𝐬𝐢𝐨𝐧? The thing you are buying next is not a better answer. It is an employee who clicks. So controls belong on the actions, not your attention, because when each action is cheap the checking quietly stops. Three quiet failures from the week: 📊 📊 A data company found its pipeline producing audience counts 𝟒𝟎% wrong for ages; every monitor checked whether the job ran, none whether it was right 📊 On the benchmark for using a computer, the new model scored 𝟒𝟏.𝟏% versus 31.4, a third better at operating the software you own 📊 On recent security vulnerabilities it found and used 𝟑𝟗%, up from 5.5; a different category of tool, not an improvement The sharp one: prompt injection now arrives inside the data your agent reads, not just what a user…