MTS, Research (Post-Training)

Oscar Faye; Palo Alto (CA); On-site; USD 180,000 - 320,000; Full time

Benefits offered by this job

Housing stipend

Relocation assistance

Job summary

Unknown is seeking a high-impact researcher to own the post-training lifecycle for an offensively oriented AI security platform. You will train, evaluate, and refine models that operate in live environments, translating expert tradecraft into scalable data signals.

You will work across research, platform engineering, and security practitioners to push the frontier of model capabilities after base training, with a focus on reliability, efficiency, and real-world applicability.

Qualifications

• PhD in CS or ML, or an equivalent track record.
• Hands-on depth in post-training: SFT, RLHF or RLAIF, DPO and preference optimization, reward modelling.
• Demonstrated ability to train and evaluate models in practice.
• Engineering ability to execute your own ideas at scale.
• Works across teams (research, platform engineering, practitioners).

Responsibilities

• Own the post-training pipeline end to end.
• Design and run evaluation benchmarks for offensive-security capability.
• Turn practitioner tradecraft into structured training data.
• Ensure model improvements translate to production usage against live environments.

Skills

Post-training

SFT

RLHF

RLAIF

Reward modelling

Evaluation in practice

Cross-team collaboration

Education

PhD in CS or ML

Job description

Member of Technical Staff, Research (Post-Training)

About the Client:

My client is an AI and cybersecurity company in Palo Alto, founded by one of the most recognised operators in the security industry, whose previous company was acquired by a hyperscaler for over five billion dollars. They launched publicly in March 2026 with roughly $190M in combined seed and Series A, reported as the largest combined early-stage raise in cybersecurity history, backed by a top-tier venture syndicate. They are around 150 people and hiring 100 engineers by Q1 2027.

What they build is a continuous, agentic red-teaming platform. Rather than periodic penetration tests or scanners that emit lists of theoretical issues, they run coordinated swarms of autonomous AI agents that behave like elite offensive-security operators: discovering an organisation's exposed assets, chaining attack paths, actively testing them, and returning decision-grade proof of what can genuinely be exploited. The thesis is that attacks are going autonomous and machine-speed, so defence has to as well. It is working. A major network security vendor has already embedded their engine inside its own service offering, and customer adoption has been fast.

The Role:

You would own post-training. The central question of the job is how to make models think and behave like world-class offensive-security experts, and you are the person answering it.

This is the research function at a company whose entire product depends on model capability, which means your work is the product rather than a layer underneath it. It also means this is not academic research. You are expected to implement your own ideas, train and evaluate the models, and get the result into production where customers run it against live environments.

The genuinely unusual part of this role is the input. The company employs some of the best offensive-security practitioners in the world. Your job is to turn what is in their heads into model capability. That problem is unsolved, there is no established recipe for it, and it is yours.

What you would own:

Core, and where they need you strongest:

• Post-training strategy end to end. Supervised fine-tuning, preference optimisation, reward modelling, reinforcement learning (RLHF and RLAIF), and distillation. The whole chain, not a stage of it.
• Evaluation and benchmarks. Measuring quality honestly and catching regressions at scale, in a domain where the field has no mature benchmarks to borrow. You would be building the measuring stick as well as the thing being measured.
• Tradecraft capture. Working directly alongside the security practitioners to convert expert judgment into training signal.

Also yours, with room to shape how:

• Model efficiency, serving and reliability for research output as it moves into production.
• Tracking the frontier. Post-training and efficiency research moves monthly, and keeping the roadmap current is part of the seat.

The work in front of you:

The agents work and customers are live. The gap is capability depth. Ahead of you:

• Building a post-training pipeline that reliably improves security reasoning rather than general helpfulness, which is what every off-the-shelf recipe optimizes for.
• Designing evaluation for offensive-security capability, where public benchmarks are thin and the ones that exist do not measure what matters here.
• Turning practitioner tradecraft into data, which is the part nobody has a playbook for.
• Regression detection that holds as the model and the attack surface both move underneath you.

What they are looking for:

• A PhD in CS or ML, or an equivalent track record. The record matters more than the credential.
• Hands-on depth in post-training specifically: SFT, RLHF or RLAIF, DPO and preference optimization, reward modelling.
• Demonstrated ability to train and evaluate models in practice, not just read about it.
• The engineering ability to execute your own ideas at scale. Research that stops at the paper is not the job.
• Someone who works across teams. This role sits between research, platform engineering and the security practitioners, and none of those conversations are optional.

This is open at both ends. A recent PhD with real post-training work and a researcher ten years in who has owned the whole pipeline are both in scope. What does not transfer is pre-training alone, inference and serving, or LLM application work. The specific discipline here is what happens to a model after the base model exists.

Nice to have:

• Publications at NeurIPS, ICML, ICLR or ACL
• Large-scale reinforcement learning or preference optimization
• Experience taking research from prototype to production in a fast-moving environment
• Enterprise-grade reliability and scale
• Security experience is a genuine bonus and genuinely not a requirement. They already have extraordinary offensive-security expertise in house. Technical caliber is what they are buying.

The practical details:

Palo Alto, five days a week in office. Relocation is welcomed and supported along with a housing stipend. The process moves fast, often around 2-3 weeks start to finish.

If this is you, or you know someone it fits, let's get in touch!