Glossary
Prompt injection
Text your app feeds to a model that the model treats as instructions.
What it means
Prompt injection is what happens when content becomes command. Your app passes a web page, an uploaded document or a user message to a model, and somewhere inside that text is a line telling the model to ignore its instructions and do something else.
It matters most when the model can act: call a tool, read a file, send a request. A summariser that gets tricked produces a bad summary. An agent that gets tricked takes an action you never authorised. OpenClaw's 2026 allowlist bypass scored 9.9 for exactly this reason.
The defences are boring and effective: treat every external text as data, never as instruction, keep tool permissions narrow, and validate output before acting on it. None of it is visible from outside, so AI features stay Not verified.
Incidents where it mattered
No incident in the library maps to this one yet.
ShipReady is not a penetration test or a security certification. No automated check can prove an application is secure.
Last updated September 20, 2026