- Deliberate misuse. Even in the status quo, terrorists are using LLMs to build missiles. As open-source models get more powerful and irresponsible people continue making abliterated versions, I expect this to only grow.
- Rogue Misaligned AI. I previously didn't think this was much of a concern, but HuggingFace, DSEWiki, and RubyHack have shown that the misaligned behavior and capacity both exist. Especially as more and more labs create more and more powerful models, the risk of a single weak link breaking grows higher.