Although there are a few ways to mitigate the risk, the only way to block it is to get AI to differentiate instructions from data, which is impossible today.
But since cloaking instructions inside data has been a thing since little Bobby Tables, and I’ve been hearing about this vector for months, it’s safe to assume someone somewhere has done this with CoPilot and Word by now.
True, itś a bad headline.
But since cloaking instructions inside data has been a thing since little Bobby Tables, and I’ve been hearing about this vector for months, it’s safe to assume someone somewhere has done this with CoPilot and Word by now.