Research Engineer, Safety & Alignment
At a glance
Mid-level AI Research role at Cognition. San Francisco · full-time · no salary stated · we estimate $200,000–$266,875.
Cognition has not published a salary for this role. The figure below is our estimate, not theirs.
aimlroles summary, based on the employer's posting.
What you'll do
- Build safety tests for agent failures such as unsafe actions, prompt injection, data theft, sandbox escape, reward hacking and misuse
- Probe Devin and its supporting systems for weaknesses, then turn findings into fixes and repeatable regression checks
- Develop training approaches with the post-training team, using reward models, preference data and constitutional methods
- Work with the agent team on permissions and oversight, including when Devin acts, seeks help or explains its choices
- Share safety work through research papers, evaluations and open methods when appropriate
What you bring
- Hands-on safety or alignment research, such as evaluation, red-teaming or interpretability work, with evidence it changed training or deployment
- Practical understanding of agent failures, including tool misuse, specification gaming, long-horizon drift or adversarial inputs
- Python and PyTorch or JAX skills, plus experience building evaluation infrastructure and running experiments at scale
- A PhD in a relevant technical field, or equivalent industry research experience
Who this fits
This suits a hands-on research engineer who wants safety work embedded in a deployed agent, and who is comfortable attacking systems and evaluating results rigorously. You’ll be a founding member of the safety team, report to its Head of Safety and work on-site in San Francisco.
From the employer
We are an applied AI lab building end-to-end software agents.
We're the makers of Devin, the first AI software engineer.
Our team is extremely talent-dense. Among our founding team, we have world-class competitive programmers, former founders, and leaders from companies at the cutting edge of AI including Scale AI, Palantir, Cursor, Waymo, Tesla, Lunchclub, Modal, Google DeepMind, and Nuro.
Building Devin is just the first step—our hardest challenges still lie ahead. If you’re excited to solve some of the world’s biggest problems and build AI that can reason on real-world tasks, apply to join us.
Role Mission
Devin writes and runs code, uses tools, takes actions in real customer systems, and operates for hours without a human in the loop. Safety here is not a research paper; it is whether an agent behaves correctly in production, millions of times a day. You will be a founding member of Cognition's safety team, reporting to the Head of Safety. You will build the evaluations and red-teaming that gate every release, develop alignment methods that shape how our models are trained, and work directly with the agent team on how Devin plans, acts, and asks for help. This is a hands-on research engineering role for someone who wants their safety work to run in the loop of a real agent, not sit in a benchmark.
What You'll Accomplish
Build agentic safety evals: Design and run evaluations for the failure modes that matter for autonomous agents: unsafe actions, prompt injection, data exfiltration, sandbox escape, reward hacking, and misuse. Make them fast enough to run on every model and product release.
Red-team Devin: Attack the agent and its harness systematically. Find the failures before customers do and turn them into fixes and regression tests.
Develop alignment methods: Work with post-training on reward modeling, preference data, constitutional approaches, and other techniques that make the model safer without making it worse at the job.
Shape the agent harness: Partner with the agent team on permissions, oversight, and escalation: when Devin should act, when it should ask, and how it should explain itself.
Publish and share: Contribute to Cognition's external safety work through papers, evals, and open methods where it makes sense.
Exceptional Candidates Have Demonstrated
Safety or alignment research experience: Hands-on work on evaluations, red-teaming, alignment, or interpretability at a frontier AI lab or in published research. You have built things that changed how a model was trained or deployed.
Agentic systems understanding: You know how agents fail in practice: tool misuse, specification gaming, long-horizon drift, adversarial inputs. You have measured it or have concrete ideas about how to.
Strong engineering fundamentals: Proficiency in Python and PyTorch (or JAX). You can build eval infrastructure, run experiments at scale, and read the training and inference code.
Empirical rigor: You design clean experiments, report results honestly, and know the difference between a real improvement and noise.
Comfort with adversarial thinking: You enjoy breaking systems and are good at it.
Relevant industry experience: Prior experience at a frontier AI lab, applied AI company, or developer tools company.
Advanced degree: PhD in Computer Science, Machine Learning, Physics, Mathematics, or a related discipline; or equivalent industry research experience.
Our Commitment
Equal Opportunity. Cognition is an equal opportunity employer. We do not discriminate on the basis of any protected characteristic under applicable law. We are committed to providing reasonable accommodations for candidates with disabilities throughout the hiring process - please let us know if you need any. Cognition also considers applicants with criminal histories consistent with the SF Fair Chance Ordinance and similar local laws.
Data Privacy and Arbitration: By submitting your application, you agree to our Global Applicant and Job Candidate Privacy Notice. For US applicants, by submitting your application, you agree to the Applicant Arbitration Agreement.