ActiveFence
ActiveFence
ActiveFence este o companie de securitate cibernetică și inteligență artificială axată pe încredere și siguranță pentru platformele online. Oferta sa completă combină cercetarea de informații, detectarea conținutului dăunător cu ajutorul inteligenței artificiale și instrumente de moderare a conținutului, ajutând platformele să gestioneze riscuri precum exploatarea copiilor, dezinformarea, discursul instigator la ură, terorismul, nuditatea, frauda și alte forme de abuz online. ActiveFence sprijină echipele de încredere și siguranță de diferite dimensiuni, contribuind la protejarea a peste trei miliarde de utilizatori din întreaga lume, în mai mult de 100 de limbi.

Senior AI Researcher – Hybrid in Israel

Lead reinforcement learning evaluation, red-teaming, and security audits for open-weight AI models. Build benchmarking and post-training pipelines for ActiveFence’s AI trust, safety, and security work.

Descriere

  • Lead post-training evaluations, red-team exercises, and reinforcement learning gym audits for open-weight models
  • Design rigorous benchmarking methodologies
  • Assess large language models against sophisticated threats, including Indirect Prompt Injection attacks
  • Build post-training evaluation pipelines that measure realistic frontier-level security capabilities
  • Run post-training experiments such as GRPO with open-weight generalist models in security-focused reinforcement learning environments
  • Examine live loss curves and rollout traces to detect reward hacking, lazy policy convergence, and poorly specified verifiers
  • Assess tasks and multi-turn environments involving tool use, web navigation, and computer use for realism, threat-model accuracy, data distribution, and dataset balance
  • Produce evaluation cards covering checkpoint performance gains, failure modes, rollout tokens and turns, and task-level success rates
  • Integrate Dockerized environments, including Harbor-format environments, into internal training frameworks
  • Improve reset and statefulness semantics, concurrency, and throughput limits

Cerințe

  • M.S. or Ph.D. in Data Science, Machine Learning, Computer Science, or equivalent practical deep learning experience
  • Hands-on experience training large-scale models with reinforcement learning algorithms such as GRPO or PPO on open-weight architectures
  • Strong knowledge of LLM vulnerabilities, red-teaming practices, and defensive alignment against Indirect Prompt Injection attacks
  • Proficiency with PyTorch, Docker containerization, and distributed training architectures
  • Ability to interpret agent rollout traces, create deterministic rubrics and verifiers, and debug complex reward-shaping issues
  • Experience with standard reinforcement learning gym formats such as Harbor is preferred
  • Experience assessing complex agentic workflows in tool-use or web-browser environments is preferred
  • Familiarity with evaluating Llama-like or Mistral-like open-weight models against adversarial workloads is preferred

Locuri de muncă similare

Wikimedia Foundation

Senior Android Software Engineer (Contract, US Remote) – Wikimedia Foundation

Wikimedia Foundation

Build and improve Android experiences for Wikipedia readers around the world. Deliver accessible, scalable mobile features for Wikimedia’s free knowledge projects.

Deschide
NETGEAR

Graduate Mobile Software Engineer, Android and iOS

NETGEAR

Join NETGEAR’s mobile team in Ireland to develop native Android and iOS apps for connected networking products. Work across testing, release automation, and cloud integrations while learning alongside experienced engineers.

Deschide
Capital One

Android Mobile Engineering Manager

Capital One

Lead Android mobile engineering at Capital One, shaping scalable experiences for more than 100 million customers. Guide mobile architecture, BFF and SDUI development, testing, and AI-enabled engineering practices.

Deschide
Capital One

Android Mobile Engineering Manager

Capital One

Lead Android architecture, client BFF and server-driven UI development, and engineering practices for mobile experiences serving more than 100 million Capital One customers. Mentor engineers and guide technical delivery.

Deschide