AI Security & Agent Red-Teaming
Three weeks attacking AI systems — your own first. Prompt injection direct and indirect, tool abuse, scope escalation, data exfiltration, and the defences that actually hold.
Join the waitlistWho should take this program?
You'll get the most from this if
- You have built an LLM or agent system
- Comfortable with Python and APIs
This is not for
- Anyone wanting general cybersecurity training
- Anyone without a system to attack
What you will build
- A documented red-team report against your own agent
- A threat model for a system with real tool access
- Hardened tool scoping with tests proving the fix
What you learn, week by week
Module 1 — Attacking your own agent
- Run direct injection against your prompts
- Plant indirect injection in retrieved content
- Document what worked and why
Module 2 — Tools as attack surface
- Escalate privilege through a poorly scoped tool
- Exfiltrate data through a legitimate-looking action
- Apply least privilege and prove it holds
Module 3 — Threat modelling and defence
- Write a threat model for real tool access
- Add confirmation for irreversible actions
- Test defences the way an attacker would
How this program handles evaluation and governance
The premise is that any content your system reads — web pages, documents, email, retrieved chunks — is untrusted input. Everything else follows from taking that seriously. You will break your own system in week one, which is considerably cheaper than learning the same lesson from a customer.
Tools and stack you will use
Fees
Request pricing
Fees depend on cohort, format, and whether this runs for an individual or a team. We'll send the full breakdown.
Get fee detailsQuestions people ask
Is this legal to practise?
You attack your own systems and our lab environments only. We are explicit about scope.
Do I need security experience?
No. You need to have built something with tool access.
Will this make my system safe?
It will make its weaknesses visible and several of them fixable. Nobody can promise more than that.