2.2 Given a set of requirements, implement security controls for AI systems.
AI controls sit in two places: on the model itself (evaluation and guardrails) and in front of it (a gateway that inspects, limits and authorises requests).
Model controls
- Model evaluation
- Testing the model against safety, accuracy and adversarial test sets before and after changes.
- Model guardrails
- Rules and classifiers around the model that block disallowed topics, sensitive data or unsafe actions in inputs and outputs.
- Prompt templates
- Fixed prompt structures that separate trusted instructions from user data, reducing injection and keeping behaviour consistent.
Gateway controls
- Prompt firewalls
- A filtering layer that inspects prompts and responses for injection attempts, secrets, personal data and policy violations.
- Rate limits
- Caps on requests per user or key over time; slow down abuse, scraping and model extraction.
- Token limits
- Maximum input and output tokens per request; controls cost and limits denial-of-wallet attacks.
- Input quotas
- Allowances on how much a user can submit over a period.
- Data size
- Limits on the size of uploads or context submitted.
- Quantity
- Limits on the number of items, files or calls.
- Modality limits
- Restricting input types (text only, no images or audio) to shrink the attack surface, since hidden instructions can ride in images or files.
- Endpoint access controls
- Authentication and authorisation on model APIs so only approved apps and users can call them.
Guardrail testing and validation
- Guardrail testing and validation
- Regularly attacking your own guardrails with jailbreak and injection test suites (red teaming) and confirming they still block what they should after every model or prompt change.
Check your understanding
Loading quiz…
Original CoreQ Academy material, mapped to objective 2.2 of the publicly available CompTIA SecAI+ CY0-001 exam objectives (Exam objectives V1, document version 1.1). Not affiliated with or endorsed by CompTIA. Contains no exam questions.