We are looking for that Individual contributor, who will slot into an existing client team already running Guardrails and AI red teaming for their foundation model suite. The role digs into edge-case vulnerabilities that campaign reports surface but don't fully explain, the ideal candidate will be able to design and train ML models as well as SLM's in the security context.

What You'll Own

• Hands-on adversarial testing across the model, the application and agentic layer, and the data pipeline: multi-turn jailbreaks and guardrail bypass, prompt injection, agent and tool-chain misuse, dangerous-capability evaluation, API abuse, and, where relevant, data poisoning, model inversion and membership inference.
• Digging deeper into edge-case findings from AI red-team campaigns, turning a flagged anomaly into a fully understood, reproducible vulnerability.
• Severity-ranked findings mapped to the OWASP Top 10 for LLM Applications, the NIST AI Risk Management Framework and its Generative AI Profile, MITRE ATLAS, and EU AI Act Article 55 expectations, with evidence and clean reproduction steps.
• Remediation guidance that's actually usable, and a retest to confirm the fixes hold.

The Human Side of It

The testing is the craft. The trust is the job. Findings only matter if the right people understand and act on them.

• Works shoulder to shoulder with the client's Guardrails and AI red-teaming team, not at a distance from them.
• Translates findings into plain language: technical depth for the engineers, a clear risk picture for anyone less hands-on with the model itself.
• Stays embedded well past the first findings, through remediation, to the retest that proves it's fixed.

Requirements

• Expert-level Python programming with deep proficiency in ML frameworks such as PyTorch, TensorFlow, and Hugging Face Transformers
• Hands-on experience fine-tuning ML models and Small Language Models (SLMs) — including techniques such as LoRA/QLoRA, PEFT, instruction tuning, and domain adaptation — for both performance and robustness objectives
• Strong foundation in ML mathematics: optimization, linear algebra, probability, and statistics
• Proven ability to design and execute adversarial attacks, including evasion (adversarial examples), data poisoning, model extraction, and membership inference
• Experience implementing defenses such as adversarial training, robust fine-tuning, input sanitization, and differential privacy
• Proficiency with adversarial ML toolkits such as Adversarial Robustness Toolbox (ART), CleverHans, and Foolbox