Avoid the most dangerous situations by making sure LLMs with untrusted input produce output that's human reviewed.

Still makes an interesting way for, say, a former employee to poison the results.