Habeas Jocus (@toowoke2joke.bsky.social)

Freedom of the Press Foundation 4 days ago

Large language models used in chatbots and elsewhere are vulnerable to malicious prompts. New research shows they may also be easily “confused” into taking user input as their own ideas. That’s right, they can be “incepted.” Read our newsletter for more.

loading . . .

The ‘s’ in LLM stands for ‘security’ Researchers bypass guardrails on open source LLMs with slight role-based language changes https://freedom.press/digisec/blog/the-s-in-llm-stands-for-security/