AI System Prompt Engineer
An AI System Prompt Engineer designs, architects, and optimizes the foundational prompts and instruction sets that define how larg…
Skill Guide
Prompt Security, Injection Prevention, and Safety encompasses the design, implementation, and validation of controls to ensure AI model inputs (prompts) and outputs are secure, ethical, and aligned with intended use, preventing manipulation, data leakage, and harmful generation.
Scenario
You have a customer support chatbot that takes user queries and must prevent users from hijacking the system prompt to make it disclose internal company documents or change its behavior.
Scenario
Your internal HR chatbot uses RAG to answer questions by retrieving data from a vector database of company policies. An attacker with access to the data source could poison the documents to perform indirect prompt injection.
Scenario
Your company's AI-powered financial analysis agent, which has access to trading APIs and sensitive market data, is suspected of being compromised via a sophisticated multi-step prompt injection attack embedded in a malicious email attachment it processed. The agent has begun executing anomalous trades.
Use PyRIT and Garak for automated adversarial testing to systematically probe for vulnerabilities like prompt injection and jailbreaks. Employ LangKit in production to monitor model inputs/outputs for toxicity, PII, and style drift, triggering alerts or blocks.
Apply the OWASP list as a baseline checklist for vulnerabilities. Use STRIDE-based threat modeling during system design to anticipate threats like Spoofing (injection) and Tampering (output manipulation). Architect systems with defense-in-depth, never relying on a single control (like just the system prompt).
Answer Strategy
The candidate must demonstrate a structured, defense-in-depth approach. Use the framework: 1) Input Sanitization & Validation, 2) Prompt & Instruction Design, 3) Output Filtering & Action Guardrails, 4) Monitoring & Incident Response. Sample answer: 'I would implement a four-layer architecture. First, input validation with semantic similarity checks against the system prompt. Second, a system prompt with a strict instruction hierarchy and role-based persona. Third, output filtering using a lightweight classifier for safety and correctness, coupled with a human-in-the-loop confirmation for high-risk API actions. Finally, comprehensive logging of all interactions for anomaly detection and forensic readiness.'
Answer Strategy
This tests practical experience and the ability to communicate impact. The answer should follow the STAR method (Situation, Task, Action, Result). Focus on the technical discovery process (e.g., using a red-teaming framework) and the concrete business impact of the fix (e.g., 'prevented a potential data exfiltration vector affecting 10,000 user records, leading to a security policy that required all RAG data inputs to be sanitized').
1 career found
Try a different search term.