AI for Security Automation
Building and attacking LLM systems, from someone who runs them in production.
Foundations
- 01
What actually breaks in deployed systems versus what makes headlines, and the OWASP LLM Top 10 read critically.
Practitioner
- 02
The vulnerability class with no clean fix: why retrieval and tool use turn every untrusted document into an instruction channel, and what mitigation actually buys you.
- 03
Hands-on with the tool-use loop, structured outputs, retries and failure modes.
- 04
Where an agent genuinely beats a script and where it is a liability. Building one that reduces an alert queue without inventing findings.
Operator
- 05
Log summarisation, rule drafting, and evaluating output you cannot manually verify at volume — including how to measure whether it works.
- 06
Voice cloning, real-time vishing agents and deepfake video — attack mechanics and defensive training design, drawn from production systems.
- 07
Sandboxing, least privilege for tools, human-in-the-loop boundaries, MCP server trust, and metering spend so a runaway loop is not a financial incident.
Researcher
- 08
Capstone: attack a deliberately vulnerable agent, then write the report. Graded against a rubric rather than a flag.