Level 10 of 14 · about 3 weeks

Safety, security and governance

Assume text from the outside world is hostile and design the agent so one bad input cannot do real damage.

Modules

1. 1

   Prompt injection

   Direct and indirect attacks, and why filters alone fail

   Core 3 parts · about 4 hours

2. 2

   Excessive agency and least privilege

   Limiting what a compromised agent can do

   Core 3 parts · about 4 hours

3. 3

   Sandboxing

   Isolating code, files and network access

   Core 3 parts · about 4 hours

4. 4

   Guardrails and validation

   Checking inputs, outputs and tool arguments

   Core 3 parts · about 4 hours

5. 5

   Approvals for risky actions

   What needs a human, and how to ask well and verifiably

   Core 3 parts · about 4 hours

6. 6

   Data privacy

   Personal data, redaction, retention and erasure in agent systems

   Core 3 parts · about 4 hours

7. 7

   Threat modelling

   Mapping an agent's attack surface and choosing what to fix first

   Common 3 parts · about 3 hours

8. 8

   Red-teaming

   Attacking your own agent systematically, and measuring the fix

   Common 3 parts · about 4 hours

9. 9

   Accountability and escalation

   Audit logs, severity, who gets paged, and how to stop an agent

   Specialist 3 parts · about 3 hours

10. 10

    Regulation basics

    EU AI Act, GDPR, US state laws, and turning rules into release gates

    Specialist 3 parts · about 3 hours