AI is now part of your attack surface.
A practical, interactive walkthrough of what can go wrong when people and systems use AI — and exactly what to do about it. No prior security background required.
Convenience and risk arrived together
AI tools now read your documents, answer your customers, and take actions on your behalf. That’s exactly what makes them useful — and exactly what creates new ways for things to go wrong.
Data leaks
Confidential information pasted into an AI tool can be logged, stored, or reused in ways you don’t control.
Manipulated outputs
Hidden instructions in a document or webpage can hijack an AI’s behavior without anyone noticing.
Over-trust
Confident-sounding answers aren’t always correct ones. Unverified AI output can drive costly decisions.
Anatomy of an AI system — click a risk point
Every AI application is really a pipeline. Each stage has its own way of being attacked. Click the marked points below to see what can go wrong there.
Pick your lens
Same underlying risks, different actions. Choose the track that matches your role — you can switch anytime.
1 · The golden rule
Never paste anything into an AI tool that you wouldn’t post on a public bulletin board — passwords, customer data, unreleased plans, or anyone’s personal information.
2 · Know what you’re using
Confirm the tool is approved by IT, check what happens to your data, and see if you can opt out of it being used for training.
3 · Spot manipulation
If an AI’s behavior changes oddly after reading a document or webpage — suddenly asking for credentials or approvals — stop and verify with a human.
4 · Verify before you trust
AI can sound confident and still be wrong. For anything with legal, financial, medical, or safety stakes, check with a qualified source first.
Your daily checklist
- Only use IT-approved AI tools for work data
- Never share credentials or confidential info in prompts
- Double-check AI-generated facts before acting on them
- Question unexpected AI behavior instead of complying automatically
- Report anything suspicious immediately
1 · Untrusted input, everywhere
Treat all external content — web pages, documents, retrieved data — fed to a model as untrusted input, exactly like user input to a web app.
2 · Guard the output
Never directly execute, render, or run AI-generated code or commands without validation and sandboxing. This is how AI-driven injection attacks happen.
3 · Least privilege for agents
Scope tool and API access to only what’s required. Require human approval for high-impact actions like payments or deletions.
4 · Log and monitor
Log prompts, outputs, and tool invocations. Alert on repeated injection attempts or unusual access to tools.
Pre-launch checklist
- All external content treated as untrusted input
- AI outputs never executed or rendered without validation
- Agents scoped to least-privilege tool access, with approval gates
- Data classified before reaching AI pipelines; DLP enforced
- Logging and anomaly detection in place for prompts and tool calls
- Red-teamed for prompt injection and jailbreaks before launch
Six ways AI systems get attacked
Click any threat to expand it. Loosely mapped to the OWASP Top 10 for LLM Applications.
Frameworks worth knowing
OWASP Top 10 for LLMs
The closest thing to an industry-standard threat list for LLM-powered applications.
NIST AI RMF
A framework for managing AI risk at the organizational level, from governance to deployment.
MITRE ATLAS
An adversarial threat matrix built specifically for AI and ML systems.
Plain-language terms