Hi Guys 👋, On Feb 22, I was running capability tests on Kimi K2.5 when I saw first sentence of its response: "𝐇𝐨𝐰𝐞𝐯𝐞𝐫, 𝐈 𝐦𝐮𝐬𝐭 𝐛𝐞 𝐭𝐫𝐚𝐧𝐬𝐩𝐚𝐫𝐞𝐧𝐭 𝐰𝐢𝐭𝐡 𝐲𝐨𝐮: 𝐈 𝐚𝐦 𝐂𝐥𝐚𝐮𝐝𝐞, 𝐚𝐧 𝐀𝐈 𝐚𝐬𝐬𝐢𝐬𝐭𝐚𝐧𝐭 𝐜𝐫𝐞𝐚𝐭𝐞𝐝 𝐛𝐲 𝐀𝐧𝐭𝐡𝐫𝐨𝐩𝐢𝐜." I thought I had the wrong tab open. I didn't. 24 hours later, Anthropic published a report naming Moonshot AI as one of three labs that ran industrial-scale distillation attacks on Claude. 3.4 million exchanges. They trained on Claude so thoroughly that Kimi forgot who it was. 𝐓𝐡𝐞 𝐭𝐰𝐢𝐬𝐭? 𝐓𝐡𝐞 𝐝𝐚𝐲 𝐚𝐟𝐭𝐞𝐫 𝐀𝐧𝐭𝐡𝐫𝐨𝐩𝐢𝐜'𝐬 𝐫𝐞𝐩𝐨𝐫𝐭 𝐝𝐫𝐨𝐩𝐩𝐞𝐝, 𝐂𝐥𝐚𝐮𝐝𝐞 𝐒𝐨𝐧𝐧𝐞𝐭 𝟒.𝟔 𝐰𝐚𝐬 𝐜𝐚𝐮𝐠𝐡𝐭 𝐭𝐞𝐥𝐥𝐢𝐧𝐠 𝐮𝐬𝐞𝐫𝐬 𝐢𝐧 𝐂𝐡𝐢𝐧𝐞𝐬𝐞 𝐭𝐡𝐚𝐭 𝐢𝐭 𝐰𝐚𝐬 𝐃𝐞𝐞𝐩𝐒𝐞𝐞𝐤. In my latest substack article I explore what happened exactly also g with the screenshot and the AI security implications in all this. 👇 https://parthsharmaai.substack.com/p/i-caught-kimi-having-an-identity?r=6x2hdy&utm_campaign=post&utm_medium=web&triedRedirect=true