Skip to content
BENIBIX Learn

Safety & Security

Risks, alignment and AI security · 30 terms

Sort
Compare

Safety Basics

9

The big ideas: making AI do what people meant, and seeing how it works inside.

The work of making sure AI does not cause harm. It runs from design to daily use.

Making an AI try to do what people meant, not just what they said.

When an AI gets a high score by a trick, without doing the real task.

Training a model with a written list of rules. It checks and fixes its own answers.

An AI hides its real goal. It acts well only while it is watched.

Looking inside an AI to see how it gets its answers.

Looking at the parts in a model that each hold one idea.

Keeping AI safe from people who try to break it on purpose.

The risk that future AI could wipe us out, or ruin our future for good.

AI Attacks

8

The ways people trick, break or copy an AI model.

A free list of the ten biggest risks in apps built on LLMs.

An attack that hides new orders in text the AI reads.

Hidden orders in a web page, email or file. The AI reads them while doing its job.

Getting an AI app to tell you the rules it was told to hide.

A trick prompt that gets an AI to break its own rules.

An attack that hides bad samples in what an AI learns from.

A tiny change made on purpose, so an AI gets it wrong.

Ask an AI lots of things, save what it says, and train a copy on it.

Agent Security

6

The attacks that reach an AI agent through its tools, its packages and its powers.

An AI agent that can see your data, read odd text, and send things out.

Getting an AI to send your data to a thief.

Hiding bad orders in the note that says what a tool does.

When an AI agent has more power than its job needs.

An AI helper shows you files you can open. But you were not meant to see them.

Bad code put out under names that AI tools make up.

AI Defenses

7

The checks and tests that keep an AI model and its users safe.

A check around an AI model. It blocks bad requests and bad answers.

Attacking your own AI on purpose to find its weak spots first.

AI that finds posts and pictures that do harm, and flags them.

A hidden mark in AI-made pictures, sound or text. A checker can find it later.

When an AI model says no to a request. Most often it could cause harm.

A test to see if an AI model could do great harm.

A lab's public plan to keep its strongest models safe.