SASH Research Scientist at Neo Research focusing on AI safety research, particularly on misalignment, loss of control, and harmful manipulation in frontier AI models. Responsible for defining research agendas and driving them to published outputs.
Responsibilities
We are well positioned for research to reach the people making decisions about these models. We have presented our work to most major Chinese frontier-model developers and are running a joint evaluations project with an AI Safety Institute. We publish our results, but also work to put them directly in front of model developers, safety institutions, and policymakers who can act on them. Our mission statement describes the research questions we currently think are most important.
Own research projects end-to-end: identify important questions, formulate hypotheses, design experiments, analyse results, and publish the work.
Develop rigorous evaluations of misalignment and loss-of-control risks, including evaluation awareness, sandbagging, dishonesty, sycophancy, and unsafe behaviour in agentic settings.
Study behaviours that are difficult to observe directly, including long-horizon failure modes and cases where models may have incentives to conceal relevant behaviour.
Investigate whether existing alignment and evaluation methods continue to work across different frontier-model ecosystems.
Work closely with research engineers to turn research ideas into robust, reproducible experiments.
Contribute to Neo’s broader research agenda and communicate findings to researchers, model developers, and safety institutions.
Qualification
Strong technical writingComfortable in an ambiguousPhD or equivalent research outputReading and writing MandarinRole logistics & benefits
Required
Track record of original research in AI safety, evaluations, or a closely adjacent field.
Deep familiarity with frontier model behaviour and elicitation methodology.
Ability to define a research agenda and drive it to concrete, published output.
Strong technical writing.
Comfortable collaborating closely with engineers.
Comfortable in an ambiguous, early-stage research environment where the questions aren't settled yet.
Preferred
PhD or equivalent research output.
Publications on dangerous capability evaluations or alignment.
Familiarity with frontier lab safety reports.
Familiarity with the EU AI Act or NIST AI RMF.
Reading and writing Mandarin.
You don't need to have worked in AI safety specifically. We're also interested in researchers from closely adjacent fields whose judgment and technical depth could transfer strongly to this work