Skip to content
Reflectionai

Member of Technical Staff - Safety

San Francisco, CA; London; New York, NYOnsite · United States
Staff · FullTime · Research
POSTED 16 NOVVERIFIED 10H AGO
Apply Ask
Role
3 days opentrue age
Posted 16 Nov
Pay
$206k–$284kno pay posted
Market p25–p75
$206k–$284k
SF Bay Area · n=1222 · 365d to 19 Nov
CompanyReflectionai →
44open roles
Opened · 30d
+0
Closed · 30d
−1
4d of snapshots · 30d window
Market
9,207open · Staff Software engineering
Opened · 7d
+7238% of open
All software engineering
75,117open
All tracked boards · live 20 Nov · n=9,207
LLMs

Our Mission

Reflection is a research lab making intelligence open and accessible for everyone to use, customize, and build on. We build open models that let anyone control their intelligence and help shape the future of AI. Our mission: make intelligence open and accessible to all.

We published our approach in the Open Safety Framework and the Beam Safety Technical Report. You will be responsible for the safety of the model.

About the Role

Conduct safety post-training on Beam: Develop and improve our SFT and RL pipeline, including policy-conditioned reasoning traces, context-aware reward models and mixed-task safety environments.

Build safeguards for open-weight deployment. Develop classifiers that are cheap to run, catch nearly all harmful cases and stay reasonably precise. These include prompted and fine-tuned classifiers and linear probes on model internals, and they cover prompts, reasoning, responses and agent trajectories. Ship them with intervention guidance that inference providers and the open-source community can actually deploy.

Build safety systems for Reflection’s API & AI Factory. Develop, train and run state-of-the-art safeguards for our API and AI Factory deployments, covering both real-time and offline monitoring and enforcement. Combine classifiers, linear probes and lighter-weight detection to find and act on misuse at scale. Turn what we learn in production into better safeguards for the whole ecosystem.

Own systemic risk evaluations. Design, run and evolve capability evaluations for biological, chemical, radiological/nuclear, cyber, harmful manipulation and loss-of-control risks. Base them on clear threat models, and use them to inform OSF risk determinations and launch decisions.

Measure safety across every harm area. Grow Iris Safety Evaluations (ISE) across our > 15 harm categories, from self-harm, cyber, and child safety to fraud and privacy, in single-turn, multi-turn and agentic settings. Track over-refusal on real-world traffic, and verify that safety holds across quantized variants and deployment settings like search and custom system prompts.

Responsible release: Validate that each model meets our risk thresholds on both system risks and safety model behavior before it ships. Work with third-party evaluators and government AI safety institutes on pre-deployment testing, and help write the safety reports and model cards that explain our decisions publicly.

Put advanced capabilities in defenders’ hands. Build defensive cyber & bio tools, such as autonomous penetration testing, and the trusted access programs that get them, and more capable models over time, safely to vetted security teams.

Lead adversarial testing. Run human and automated red teaming across misuse, jailbreaks, prompt injection and alignment failures. Build attacker agents that adapt to the model, so we test against dynamic attacks rather than static datasets, and feed what we find back into training and safeguards.

Run Safety evaluations grounded in the real world. Design and run human uplift studies, from uplift trials to wet-lab studies, and live-deployment evaluations like our partnership with Andon Labs, where Beam runs a real business, to measure what our models can actually do in the real world.

Own the safety policies. The harm policies, red lines and constitution, plus working with safety subject-matter experts on data and grading.

Push the safety & alignment research frontier. Work on problems that matter most for open weights: resistance to abliteration and fine-tuning attacks, safety that holds up inside the model, monitorability, evaluation awareness and agentic security. Publish the results and release open tools the community can build on.

About You

Graduate degree (MS or PhD) in Computer Science, Machine Learning, or related discipline, or equivalent practical experience in AI Safety.

Deep technical understanding of LLM safety, including adversarial attacks, red-teaming methodologies, post-training, evaluations, and interpretability.

Strong software engineering capabilities with experience building automated evaluation pipelines or large-scale ML systems.

Source · ReflectionaiFirst seen 2 Nov2 editsOriginal ↗