Social engineering has limits and each individual has unique vulnerabilities. It's not possible to call in and speak a single sentence compelling any agent who hears it to immediately burn the office building down.
Some human vulnerabilities are surprisingly common. A lot of scammers follow scripts and formula. Of course, to coax arson would be difficult. But life-devastating incidents like emptying out the entire bank account, leaking secrets, causing self-harm, etc. are not unheard of.
Maybe eventually they will be less suspectable to social engineering but I don't have that confidence yet
Is it still social engineering if you're talking to an AI?