Adversarial Machine Learning Engineer - Red Teaming

Role Overview

Join a dedicated client team focused on Guardrails and adversarial testing for foundation AI models. This individual contributor role involves investigating subtle vulnerabilities flagged by AI red-teaming campaigns, designing and training machine learning and small language models within a security framework.

Key Responsibilities

• Conduct hands-on adversarial assessments spanning model layers, application and agentic components, and data pipelines. This includes multi-turn jailbreaks, guardrail bypass attempts, prompt injection, misuse of agents and tool chains, assessment of dangerous capabilities, API exploitation, data poisoning, model inversion, and membership inference.
• Analyze complex edge-case anomalies reported during AI red-team campaigns to convert them into reproducible and well-understood security vulnerabilities.
• Classify and prioritize findings according to standards such as OWASP Top 10 for LLM Applications, NIST AI Risk Management Framework (including its Generative AI Profile), MITRE ATLAS, and European AI Act Article 55, providing evidence and clear reproduction steps.
• Deliver practical remediation advice and verify fixes through thorough retesting.
• Collaborate closely onsite with the client’s Guardrails and AI red-teaming teams rather than working remotely or in isolation.
• Translate technical findings into understandable language suitable for engineers and stakeholders with varied technical expertise.
• Remain engaged through the remediation cycle to validate implemented corrections.

Candidate Requirements

• Advanced Python programming capabilities combined with expertise in ML frameworks such as PyTorch, TensorFlow, and Hugging Face Transformers.
• Experience fine-tuning machine learning models and small language models using methods like LoRA/QLoRA, PEFT, instruction tuning, and domain adaptation, balancing model performance and robustness.
• Solid grasp of machine learning mathematics including optimization techniques, linear algebra, probability, and statistics.
• Proven skill in performing adversarial attacks including evasion, data poisoning, model extraction, and membership inference.
• Hands-on implementation of defense mechanisms like adversarial training, robust fine-tuning, input sanitization, and differential privacy.
• Familiarity with adversarial machine learning tools such as Adversarial Robustness Toolbox (ART), CleverHans, and Foolbox.
• Experience conducting red-teaming exercises targeting AI and LLM systems, including prompt injection and jailbreak testing, with a focus on safety and alignment.
• Competence in evaluating and benchmarking the robustness, security, and safety of models pre- and post-fine-tuning.
• Understanding of MLOps best practices, encompassing model version control, experiment logging, and secure deployment pipelines.
• Strong threat modeling abilities combined with an attacker’s perspective and capacity to articulate risks effectively to both technical and non-technical audiences.
• Awareness of current research trends in adversarial machine learning and generative AI security.

Additional Information

• Work is fully remote within Canada, prioritizing output and delivery over physical presence.
• Opportunities exist for career advancement towards staff and principal technical roles.
• Comprehensive support and accountability provided by C-Serv throughout hiring and beyond.
• The company values empathy, integrity, collaboration, growth, and is woman-owned.