We're using small language models to detect prompt injection. Not too cool, but at least we can publish some AI-related stuff on the internet without a huge bill.