Applied Data Scientist — Intent Drift Detection
Backslash Security · Realtime AI-agent protection
About Backslash
We are builders who want to secure the next generation of software development from the inside out.
Software development has entered a new era defined by vibe-coding, MCP servers, and AI agents. While these tools allow developers to build faster than ever, they also introduce a completely new frontier of security challenges that exist directly on the engineer's machine.
Backslash is building the security layer for this new reality. We provide the essential guardrails and visibility needed to secure AI-driven workflows. By actively protecting developer computers and monitoring IDE-based AI interactions, we ensure the future of coding is as secure as it is fast.
What you'll do
AI coding agents run with enormous privilege on developer machines. When untrusted content steers one off its task, classical security sees only the final action — never that the agent's intent drifted. Detecting that drift, in realtime, on-device, is an open problem. No one has solved it. You will.
Build the Dataset: Design and generate the data that defines what "drift" looks like, and co-design the pipeline that creates it.
Train Small, Ship Local: Fine-tune, distill, and shrink the models that detect drift in realtime — small enough to run fully on-device, no cloud.
Shape the Architecture: Decide what's a model, what's a rule, and what's a gate. The detector's design is your call.
Measure What Matters: Build the evaluation harness — realtime detection at very low false-positive rates on a developer's machine is the product.
You own intent-drift analysis as a problem, not a ticket.
Requirements
Applied Data Science End-to-End: 3–5 years turning ambiguous problems into working models — problem framing, data collection and labeling strategy, feature/representation design, training, and evaluation. You've owned a dataset and the model that consumed it.
Small-Model Fluency: Comfortable fine-tuning small open models locally — DeBERTa-class encoders, ≤8B decoders, LoRA, quantization, single-GPU workflows.
Applied-Research Mindset: You approach an open research problem with passion and a practical, ship-it attitude.
Independence: You want an impactful, strategic problem in LLM security — not a ticket queue.
Startup DNA: Can-do, fast, a true partner. Non-negotiable.
Local-first by design. Python, PyTorch, HuggingFace, ≤8B models.
Nice to have
Security Background: Prompt injection, MITRE ATLAS, agent security — or the hunger to learn it fast.
Adversarial ML: Experience with LLM evaluation, red-teaming, or building adversarial datasets.
Endpoint Software: Familiarity with software that runs locally on developer machines with a low resource footprint.