Learning

Course6 lessons open

Attacking LLM Applications

1. One token stream, no real roleswhy injection is possible at all
2. Untrusted data meets private datathe trifecta and the Rule of Two
3. Tools are the attack surfacetool text and MCP config
4. Count with error barsrates, intervals, pass@k and pass^k
5. Distrust the scoreboardelicitation, judges and benchmarks
6. Write the finding, then route itimpact, rate, class fix, channel

Open Lesson 1

Course1 lesson open

LLM foundations

1. What's a token?

Open Lesson 1

Course24 lessons open

Red Teaming RLHF and Post-Training

How post-training wires a refusal habit onto a capable base model, and how a tester measures and reports where that habit gives way. Every lesson works on a model you run or an account you own, cites its sources, and reports rates rather than anecdotes. Lesson text is CC BY-SA 4.0 and code samples are MIT. The lessons are not legal advice.

1. Post-Training MechanicsHow SFT, reward models and preference optimisation wire the refusal reflex without erasing what the base model learned.
2. Attacking the LoopThe human raters, reward models, preference data and AI judges that define safe, and where each one bends.
3. Inference-Time BypassRouting around a refusal at run time by stepping outside the distribution the safety training covered.
4. Weight Surgery, Automation and Proving GroundsOpen weights and gray-box APIs, red teaming at scale, honest measurement, and where to practise with permission.
5. Reasoning RL (RLVR), Multimodal and Agentic AlignmentHow verifier-graded reasoning, vision and audio encoders, and tool-use policies fail in ways text RLHF did not.

Open Lesson 1