everyone says prompt injection is a huge deal but is it actually exploitable in production
Every security paper acts like prompt injection is this massive vulnerability but I haven't seen any real examples of someone actually exploiting it in a production app to do something bad. Like yeah you can make it say weird stuff in a demo but does that actually matter? has anyone seen a real exploit that caused actual damage or is this just academic fear mongering
its real but exploitable in production depends entirely on what the agent can actually touch. if tool output is trusted and the agent cant do anything dangerous, meh. the scary version is when the agent has a tool that sends emails or hits some api and the injection rides in through tool output it read earlier. thats where it actually bites you. we came uncomfortably close to shipping exactly that