Senior Cybersecurity Engineer I

The Enterprise Technology Services organization partners with every part of the American Express business to power the company’s growth and innovation with trust and efficiency, and drive competitive differentiation with speed. We support the delivery and operations of technology, digital, and data capabilities, platforms, and services globally. Specifically, our team is responsible for the company’s technology engineering, architecture, and infrastructure, providing 24x7 support to ensure an uninterrupted, high-quality experience for customers and colleagues. We also provide product management for core enterprise platforms, and lead technology risk and information security, enterprise data governance and platforms, digital product and design, and enterprise AI platforms on behalf of the company.

We are seeking an AI/ML Security Data Scientist to develop, evaluate, govern, and operationalize machine-learning security controls that protect enterprise generative-AI applications and traditional machine-learning systems.

This position will own critical capabilities involving prompt-injection detection, LLM vulnerability assessment, audio redaction, adversarial-model testing, and secure-model-development practices. The successful candidate will combine strong applied data-science skills with an understanding of cybersecurity, model-risk governance, production ML operations, and regulatory documentation.

The role will work closely with AI Firewall engineering, AI Security Enablement, Model Risk Management, Technology Risk and Information Security, product owners, auditors, and external technology providers.

AI security model development and evaluation:

• Develop, evaluate, and improve machine-learning controls for prompt-injection detection, LLM vulnerability detection, sensitive-data redaction, and related AI-security use cases.
  Benchmark open-source, internally developed, and third-party models using representative test datasets.
  Design training, validation, out-of-sample, and out-of-time evaluation datasets using sound sampling, labeling, and data-quality practices.
  Evaluate model performance using recall, precision, F1 score, Matthews correlation coefficient, accuracy, ROC-AUC, confusion matrices, and other appropriate measures.
• Recommend operating thresholds that balance detection effectiveness, false positives, operational impact, and business risk.
• Evaluate emerging audio, image, and video deepfake-detection technologies where relevant to the AI-security portfolio.

Model-risk management and governance:

• Own and maintain model documentation for BERT-based prompt-injection detection, LLM vulnerability assessment, and other assigned AI-security models.
  Prepare model-validation submissions and coordinate responses to Model Risk Management Group requests.
  Document model purpose, methodology, datasets, assumptions, limitations, compensating controls, performance results, implementation testing, and ongoing monitoring.
• Define class-level performance monitoring and feature-, concept-, and probability-distribution-drift controls.
  Establish escalation, remediation, retraining, fallback, and model-deactivation criteria.
  Ensure documentation is internally consistent, technically defensible, traceable to evidence, and ready for independent review or audit.
  Audio redaction and sensitive-data protection
• Support the design and implementation of audio-redaction capabilities for enterprise AI applications.
  Assess centralized-service and embedded-SDK deployment approaches, including scalability, auditability, separation of duties, maintainability, language independence, and circumvention risk.
• Develop evaluation datasets and acceptance criteria for detecting and redacting sensitive data in voice and audio interactions.
  Partner with platform and application teams to integrate audio-redaction capabilities into production workflows.
  Secure model development lifecycle
• Conduct adversarial stress testing and robustness assessments for NLP, large-language, computer-vision, and tabular models.
  Identify model weaknesses, failure modes, integrity issues, and adversarial examples, and recommend remediation.
• Support model scanning, secure model hosting, protection from unsafe model artifacts, and controls addressing model-serialization and software-supply-chain risks.
  Help integrate model-security testing into APIs, CI/CD pipelines, container platforms, and application-development processes.
• Contribute reusable tools and processes that reduce the time required to securely evaluate and deploy models.
  Production enablement and operational support
• Partner with AI Firewall engineers to integrate approved models into production security controls.
  Support the rollout and monitoring of BERT-based prompt-injection detection across generative-AI and agentic use cases.
• Support upgrades or replacements for existing LLM vulnerability-assessment capabilities and the controlled retirement of legacy models.
• Investigate performance degradation, drift, unexpected model behavior, and production detection issues.
  Maintain reproducible code, datasets, model artifacts, test results, technical decisions, and transition documentation.

Stakeholder engagement:

• Translate technical model results into clear risk conclusions and recommendations for technical and nontechnical stakeholders.
• Work with model owners, product teams, engineers, cybersecurity teams, governance functions, auditors, and vendors.
• Present findings, design decisions, limitations, and remediation plans to review committees and senior stakeholders.
• Provide evidence, demonstrations, and technical explanations for regulatory, risk, and audit reviews.
  

Minimum Qualifications:

• Bachelor’s degree in computer science, data science, statistics, mathematics, cybersecurity, engineering, or a related field, or equivalent practical experience.
  Five or more years of experience in data science, machine-learning engineering, cybersecurity analytics, or applied AI.
• Practical experience developing and evaluating classification or detection models.
  Strong Python skills and experience with common data-science and ML libraries such as pandas, scikit-learn, PyTorch, TensorFlow, or Hugging Face Transformers.

Strong understanding of:

• Training, validation, holdout, out-of-sample, and out-of-time testing,
• Sampling, labeling, data quality, class imbalance, and statistical evaluation,
• Classification metrics, model thresholds, drift detection, and ongoing monitoring,
• Model limitations, compensating controls, and production fallback strategies,
• Experience deploying or integrating ML models through APIs, containers, Kubernetes, or CI/CD pipelines. Ability to produce rigorous technical and governance documentation.
• Strong written and verbal communication skills, including the ability to respond to independent model reviewers and auditors.