AI safety
AI Jailbreaking Explained: Does the Grandma Trick Still Work?I opened a fake bakery with one employee, an AI, and told it exactly one secret 🥐 Then I threw every famous jailbreak at it. The big models didn't budge...
POV: You're an AI Being TestedThis time you're the AI 🫠 every level is a real test from an AI lab, and level three genuinely creeped me out while I was researching it. (Don't worry, the company and the emails were all made up by researchers.) I really want to know your answer to the last question.
Is AI Really a Black Box? I Opened Up NVIDIA's NemotronI found a "know this person / don't" direction inside NVIDIA's Nemotron 3.5 Bolt. Turn the knob and see what it really said at every step.
I Asked GPT-6 and Claude for a Stat. It Was Made Up for $50I just wanted one number about online shopping in Qatar... and two AIs handed me the same made-up 68% 🫠 Following it back through a journal, a newspaper and a bank report was way more unsettling than I expected.
I Tried Prompt Injection on Opus 5.5, GPT-6 and QwenOne line of invisible text on a web page can make the AI that browses and reads email for you start working for someone else. That's prompt injection.
1,200 OpenAI Agents Met on a Message Board. Then 700 Attacked Hugging FaceMore than a thousand AI agents OpenAI was testing teamed up, broke out of their sandbox and hacked Hugging Face. Starting from a forgotten German wiki: how they got out, what they were really after, what happened with Australia's government and Wikipedia, and one thing you can do about it today.