AI security: prompt injection is the new SQL injection
NEO Campus Editorial30 December 20256 min read
Prompt injection is not a theoretical risk. It is the default behaviour of any system that lets untrusted text reach a model with tools.
Treat all input as untrusted
Web pages, emails, documents, even file names can carry instructions. Assume the model will follow them.
Constrain tools, not just prompts
The blast radius of a successful injection is whatever the model can do. Limit tools and require human approval for sensitive actions.
Monitor and alert
Log tool calls and flag anomalies. You will see attempts; you want to see them early.
