Glossary · Safety and data
noun
An attempt to hijack an AI system by hiding instructions in content it reads, such as a web page, email or document. It is guarded against with limited permissions and human approval for actions.
In practice
Prompt injection is a risk whenever AI reads content from outside the business, such as emails, web pages or uploaded documents, and can take actions. Hidden instructions in that content might try to make it do something it should not. The defences are practical: give AI only the permissions it needs, keep a person approving actions and treat external content as untrusted.
Limit what the AI can access and do, require human approval for actions such as sending, posting or paying, separate trusted instructions from untrusted content and monitor what automated tools do.
From AI curious to AI capable
Start with an audit. You will know exactly where AI fits, what to do first and what it should give back. Your report looks something like this.
Prepared for
your business
What Claude can really do
Contents
Every opportunity mapped, scored and turned into a costed Claude roadmap.