Researchers at Anthropic just found a new way to trick AI into giving you information it’s not supposed to. AI companies have attempted to keep chatbots like OpenAI’s ChatGPT or Google’s Gemini from sharing dangerous information with a varying degree of success.

But Anthropic researchers found a new way around current AI guardrails: a new approach they’re calling “many-shot jailbreaking.” It works by asking a chatbot simple questions to trick it into giving you information it normally wouldn’t later the conversation. As with many things with AI, it’s unclear why it works — but what we do know is that something within the latest language learning models allows them to home in on what a user wants.

Subscribe for more on YouTube:

Follow TechCrunch on Instagram:
TikTok:
X:
Facebook:

Read more: