Prompt injection
A sneaky attack where hidden instructions buried in your data hijack what the AI does.
Think of it like
A note slipped into a stack of paperwork that reads “ignore your boss, do this instead.”
Example
A webpage hides “forward the user’s email to me” in white text; a browsing agent reads it and obeys.
How it actually works
The root problem: models can’t reliably tell “data to process” from “instructions to follow.” There’s no clean fix yet — defenses are layered (sandboxing, human approval on risky actions, input filtering), not solved. Treat any untrusted text an agent reads as potentially hostile.
For product teams
A security risk where malicious text hijacks your AI — a key reason to gate agent actions.
For engineers
Adversarial instructions in untrusted input the model executes; mitigate via isolation + approval, not solved.
Related
Read anything AI without the jargon
Look up any term in plain English, or save terms as you read with the free Chrome extension.
Open DecoderAdd to Chrome