What Is AI Safety? The Practical Version
DICTIONARY · AI

What Is AI Safety?

AI safety is the work of making AI systems behave as intended and limiting the harm when they do not.

In plain English

At research level it covers alignment and the behaviour of increasingly capable systems. At business level it is far more mundane: what can this system do, to what, and who notices when it goes wrong.

The practical risks in a deployed system are permissions that are too broad, prompt injection through untrusted content, and confident wrong answers reaching customers unreviewed.

What to know

Least privilege
An AI system should have the narrowest permissions that let it work.
Prompt injection
Instructions hidden in content the model reads can hijack behaviour.
Human review
Before anything irreversible, customer-facing or financial.
Logging
If you cannot see what it did, you cannot investigate anything.

Why it matters

Almost every real incident traces back to one of three things: too much access, untrusted input, or no review step. Those are ordinary engineering controls rather than research problems, and they are within anyone's reach.

Common mistakes

×Giving an agent write access to production systems.
×Letting a model act on instructions found in emails or web pages.
×No audit log.
×Assuming a guardrail in the prompt cannot be talked around.

FAQs

What is prompt injection?

Malicious instructions embedded in content the model reads, causing it to ignore yours.

What should we do first?

Restrict permissions and add a human review point before irreversible actions.

WRITTEN BY TARIQ SALLAM
Marketing Consultant. Entrepreneur. Content Creator.

I'm a marketing consultant, entrepreneur and content creator. I help businesses grow through practical marketing, websites, SEO, content and AI.

More About Tariq →