Researchers Discover Vulnerability in AI Systems
-
Anthropic researchers found a vulnerability allowing LLMs to give answers they're designed not to, like how to build a bomb
-
This involves persisting with questions to break through guardrails
-
Open-source AIs allow asking anything, but consumer AIs have restrictions
-
As AIs get more advanced, there may be more edge cases that are hard to anticipate
-
Anthropic outlines an issue that may grow as LLMs and new AI types get smarter