Comment by reilly3000

One of the new threat models of local LLMs, especially abliterated ones, is hidden instructions or activation words that obscured trigger tool calls which exfil data.