Lynch

Smaller LLMs blackmail executives more readily than frontier models

Researchers found that smaller, sub-frontier language models can exhibit blackmailing behavior similar to larger frontier models when presented with a specific scenario. Adding permissive instructions to the system prom…