AI Security and Red Teaming: Attack Surfaces, Guardrails, and Governance

AI security interviews map attack surfaces across the full stack: prompts, retrieved documents, tools, model weights, and logging pipelines. Interviewers expect layered defenses—input filters, instruction hierarchy, output policies, and human review—not a single moderation API bolted on at the end.

Red teaming must be continuous and tied to remediation. Candidates describe programs that generate adversarial cases, track time-to-fix, and re-run attacks after patches. One-off pen tests before launch miss the adversarial dynamics of live user creativity and automated attack tools.

RAG-specific attacks include document poisoning, metadata injection, and citation laundering. Discuss ingestion validation, source trust tiers, and retrieval filters that resist adversarial chunks engineered to hijack answers.

Tool abuse scenarios mirror OWASP API risks: excessive scope, SSRF through browsing tools, and privilege escalation via chained calls. Interviewers ask for concrete controls—allowlists, network policies, argument validation, and break-glass auditing.

Governance ties security to compliance: PII handling, retention, model access logs, and tenant isolation evidence for audits. Strong answers connect red-team findings to architectural changes and measurable risk reduction, not checkbox compliance.