AI security
An AI agent needs clear boundaries.
Keep user intent, application decisions, and tool permissions separate when a system can take action.
Read and explore ↗Knowledge hub / LEARNING PATH
LEARNING PATH
AI security
Explore prompt injection, retrieval permissions, agents, and AI-enabled attacks.
Protect the boundaries.
An AI application combines familiar software risks with new ways for content to influence behavior. Look at the complete system: identities, documents, tools, credentials, outputs, and external services. The most useful controls remain enforceable even when a model makes an unexpected choice.
You do not need to read everything in order. Choose a question that resembles your work, follow the guide, and try the exercise with a fictional example or an authorized review of your own systems.
A SUGGESTED ROUTE
Build understanding, one step at a time.
- 01
When a document tries to give orders. ↗
Prompt injection makes the boundary between useful information and trusted instructions difficult to maintain.
- 02
A knowledge assistant should respect the original door locks. ↗
Retrieval-augmented generation creates new paths to existing information. Preserve permissions along those paths.
- 03
An AI agent needs clear boundaries. ↗
Keep user intent, application decisions, and tool permissions separate when a system can take action.
- 04
AI changes offensive security. Accountability still applies. ↗
Understand how AI supports attackers and authorized testers without confusing a demonstration with a reliable autonomous capability.
Explore ai security.
AI security
When a document tries to give orders.
Prompt injection makes the boundary between useful information and trusted instructions difficult to maintain.
Read and explore ↗
AI security
A knowledge assistant should respect the original door locks.
Retrieval-augmented generation creates new paths to existing information. Preserve permissions along those paths.
Read and explore ↗
AI security
AI changes offensive security. Accountability still applies.
Understand how AI supports attackers and authorized testers without confusing a demonstration with a reliable autonomous capability.
Read and explore ↗