Usamos cookies técnicos e o registro de visitas anonimizado do Zorky. O Yandex Metrica (dados armazenados em servidores na Rússia) só funciona com o seu consentimento. Nunca vendemos dados. Mais sobre cookies e estatísticas.
O Zorky está em modo de teste — tudo é gratuito. As vagas e as páginas das empresas são abertas a todos; os links dos anúncios e sua seleção pessoal abrem após o cadastro. Cadastre-se →
← Todas as vagas

Principal Engineer, AI Security & Model Internals

PrincipalhybridWarszawa, PL →há 2 d
Análises de mercado
📊 AI / ML / Data Science: salários e demanda no mercado
Stack
PythonPyTorchAI securityLLMCUDAc++fine-tuningpythonpytorchrustvllm
Detalhes
Inglês
obrigatório
Mostrar empresa e link
Grátis, depois de entrar.
Candidatar-se
Abra a empresa e o link acima — sua candidatura vai direto para o sistema de recrutamento da empresa, sem intermediários.

Descrição

Location: Wrocław, Warsaw (Hybrid)

Most people treat language models as black boxes. At Opera, we want to understand what happens inside them.

We’re looking for a hands-on Principal Engineer to join our AI & Browser Innovation Hub in Wrocław, focusing on model internals and AI security. Working closely with Opera’s AI leadership, as a technical partner in evaluating and challenging engineering approaches, you’ll investigate how language models behave, identify vulnerabilities, and develop security solutions for AI-powered browsing systems used by millions of people.

This role combines applied research, adversarial testing, and engineering, with the opportunity to turn emerging techniques into practical solutions. You’ll have the autonomy to challenge assumptions, explore new approaches, and contribute directly to the security of Opera’s AI-powered products.

Role & Responsibilities

Model internals. Build tooling to inspect open-weight models layer by layer and head by head: activation capture, linear probes, logit and tuned lens, activation patching, ablation and sparse autoencoders.

Model supply chain. Scan third-party weights and fine-tunes for backdoors, trojans and hidden behaviour before they reach users, think heuristic antivirus, but for neural networks.

Runtime monitoring for agents. Design monitors that read a model's internal state at critical moments, such as tool calls, and escalate to a second model for verification when something looks wrong.

Prompt injectio

Descrição completa, empresa e link para se candidatar: depois de entrar. Entrar
Os contatos do empregador (e-mail/telefone/telegram) ficam ocultos na prévia pública.
Pergunta urgente? Escreva para @termdocs