Week 7 · Phase 3: Attacks on AI

OWASP Top 10 for LLMs & agent risks

The checklist every AI security person uses, plus the new agent risks.

Time this week: 7–9 hours

Goal for this week
By Sunday you can name all 10 OWASP LLM risks (2025), give an example of each, and stop an over-powered AI agent in the lab.
Why it matters for AI security
When you review any AI app at work, this list is your checklist. Interviewers and both AI certs expect you to know it well.

Start here: this week's labs

Agent Hijack Lab: excessive agency
An email assistant with too much power reads a booby-trapped email. Watch the attack, then fix permissions and approvals until it fails.
Open lab

An over-powered agent (LLM06 Excessive Agency)

Malicious emailhidden instructions
AI email agentreads inbox
Toolsread, send, forward, delete
No approvalacts on its own
Data leavesforwarded to attacker

Key ideas

LLM01 Prompt Injection

Inputs change the model's behaviour in ways the developer didn't want.

LLM02 Sensitive Information Disclosure

The model leaks personal data, secrets or business info.

LLM03 Supply Chain

Risky third-party models, datasets, plugins or libraries.

LLM04 Data and Model Poisoning

Bad data in training or fine-tuning that plants backdoors or bias.

LLM05 Improper Output Handling

The app trusts model output blindly, e.g. runs it as code or shows it as HTML (XSS).

LLM06 Excessive Agency

The AI has too many tools, too much permission, or acts without a human check.

LLM07 System Prompt Leakage

Hidden instructions (and any secrets in them) get exposed.

LLM08 Vector and Embedding Weaknesses

Problems in RAG: poisoned docs, or users retrieving data they shouldn't see.

LLM09 Misinformation

False but convincing output (hallucinations) that people trust.

LLM10 Unbounded Consumption

No limits on use: huge bills, denial of service, or model theft by mass queries.

Day-by-day plan

Tick each day when you finish it. Your ticks are saved in this browser.

Mini game: OWASP LLM Top 10 sorting game

Practice outside this course (free)

OWASP Top 10 for LLM Applications 2025
The official list with examples and fixes.
PortSwigger: Web LLM attacks labs
Free labs, e.g. 'Exploiting LLM APIs with excessive agency' and 'Indirect prompt injection'.
OWASP Top 10 for Agentic Applications (2026)
Released Dec 2025. Risks specific to AI agents.

These links need the internet. Everything else on this page works offline.

Exercises & worksheet

Exercise deck (PPTX) Worksheet (DOCX)

Weekly quiz

Finished the week?
Mark it done to fill your progress bar on the course home.
← Week 6: Prompt injection & jailbreaks Week 8: MITRE ATLAS & AI threat modeling →