The LLMs are going to get even more capable. I would use a closed weights SOTA for security analysis of my software. But sometimes those SOTA models refuse to help! Next best thing is an open weights model on MapleAI (nprofile…uh5t) or self hosted or hosted on a cheap platform like vast.ai. And if they refuse as well, try uncensored (abliterated) models.
If the software should be kept private, you need to invest in running local LLMs..
As new models are released it is going to be a good practice to automate these security checks.. Red teaming uncensored LLMs to fight yourself and try to find exploits... Weird times.
