Вакансия из Hirify Global, списка международных tech-компаний Для мэтча и отклика нужен Plus
Пожаловаться
Зарплата и рынок
ЗП не указанаразвернуть
На похожих ролях по рынку
$65к$77кмедиана$108к
29 в выборке
На СНГ рынке: $4.8к/мес ($3.8к - $6.7к)
91
Отличная вакансия
развернуть
Чётко описанная роль principal-уровня в продуктовой компании сочетает передовые исследования в AI-безопасности с прикладной инженерией для массового продукта и даёт заметную техническую самостоятельность.
Передовой стек
Чёткая зона ответственности
Влияние на продукт
?Гибридный формат
Мэтч & Сопровод
Для мэтча с этой вакансией нужен Plus
Подключить Plus
Описание вакансии
Текст:
Hirify AI/Оригинал
TL;DR
Principal Engineer (AI Security/Model Internals): Building model-inspection tooling, supply-chain scanners, runtime monitors, prompt-injection defences, and secure fine-tuning pipelines for Opera’s AI-powered browsing systems with an accent on mechanistic interpretability, adversarial testing, and on-device inference. Focus on inspecting transformer internals, detecting backdoors and data poisoning, evaluating model behaviour, and converting security research into production-ready safeguards.
Location: Hybrid in Wrocław or Warsaw, Poland
Company
hirify.global develops AI-powered browser products used by millions of people and hirify.globaltes an AI & Browser Innovation Hub in Poland.
What you will do
Build tooling to inspect open-weight models layer by layer using activation capture, probes, lens techniques, activation patching, ablation, and sparse autoencoders.
Scan third-party model weights and fine-tunes for backdoors, trojans, and hidden behaviour before deployment.
Design runtime monitors for AI agents that inspect internal states during critical actions and escalate suspicious behaviour for verification.
Red-team agentic browsing systems and develop architectural defences against prompt injection and untrusted web content.
Build and evaluate LoRA-based secure fine-tuning pipelines resistant to data poisoning, memorisation, leakage, and safety regressions.
Reproduce promising research quickly and turn validated techniques into monitors, checks, and open-source tools suitable for on-device inference.
Requirements
Strong engineering experience with Python and PyTorch, including production systems rather than only notebooks.
Hands-on experience with transformer internals, including the residual stream and model forward-pass hooks using TransformerLens, nnsight, or raw hooks.
Practical fine-tuning experience with LoRA, QLoRA, or PEFT, including evaluation of trained models.
Demonstrable adversarial experience such as red-teaming, vulnerability research, reverse engineering, CTFs, or jailbreak work.
Ability to take research from a paper to a working prototype quickly and distinguish genuine findings from artefacts.
Clear written English and confidence to challenge technical ideas, including those of a manager.
Nice to have
Mechanistic interpretability experience with sparse autoencoders, circuit analysis, probing, or related published or open-source work.
ML security research involving backdoor detection, data poisoning, model extraction, or membership inference.
Experience with inference runtimes and performance, including llama.cpp, vLLM, MLX, CUDA or Metal kernels, and quantisation.
Browser security, sandboxing, or agent-framework experience.
C++ or Rust experience for integrating monitors into native code paths.
Experience taking apart systems such as game trainers, memory editors, emulators, or firmware.
Culture & Benefits
Diverse, inclusive, and collaborative team of experienced professionals.
Flat structure with short decision-making paths and an environment that values creativity.
Support for ongoing skill development in a friendly and empowering workplace.
Applications are reviewed on a rolling basis.
Hiring process
Submit a CV in English.
Consult the FAQ for recruitment process, remote-work, and benefits information.
Будьте осторожны
Если работодатель просит войти в их систему, используя iCloud/Google, прислать код/пароль, запустить код/ПО, не делайте этого. Это мошенники. Обязательно жмите «Пожаловаться» или пишите в поддержку. Подробнее в гайде →