AI Security: AI Gateway, AI Guardrails, Workforce AI Security
What the domain tests
How you control traffic to and from large language models, inspect what goes in and comes out, and govern employee use of AI tools. The exam expects you to distinguish the three control surfaces and to know which risk each one actually reduces.
Core rules
- AI Gateway. A single egress control point for LLM and agent traffic: provider credentials are vaulted so clients never hold them, model routing and failover handle rate limits and outages, per-user token quotas control cost and abuse, and prompt/response logging supports audit and data-residency requirements.
- AI Guardrails. Inline inspection of prompts and responses. Inbound: prompt injection, jailbreak patterns, encoded or obfuscated payloads, policy-violating requests. Outbound: PII and secret redaction, harmful content, grounding checks for hallucination. For agents, constrain which tools and actions are permitted and require approval for high-impact operations.
- Workforce AI Security. Discover unsanctioned AI SaaS, apply DLP to data pasted into those tools, enforce tenant restrictions, and coach users toward approved alternatives.
- Threat mapping. The OWASP LLM risks include prompt injection (direct and indirect through retrieved documents), insecure output handling, model denial of service and supply-chain risk. Indirect injection via RAG content is the one teams most often miss.
Worked example
An internal chat application calls an external model. The client talks only to the AI Gateway, which injects the provider key, applies a per-user token quota, and fails over to a secondary model on HTTP 429. A guardrail scans the prompt for injection patterns and redacts account numbers, then scans the response for secrets before returning it. Logs are retained with sensitive fields masked. Reasoning: the gateway solves credential exposure, cost and availability; the guardrail solves content risk; neither alone covers the other.
Traps that cost marks
Relying solely on the model vendor safety filter. Treating prompt injection as a regex problem. Ignoring indirect injection arriving through retrieved documents or web pages. Giving agents unrestricted tool access. Having no visibility into employee use of external AI services, so sensitive data leaves with no log at all.