Senior Research Scientist/Engineer, Deception

We are looking for an experienced research scientist/engineer to partner with the Research Lead of the Deception team. You’ll lead research projects from initial proposal to implementation, with scope to contribute across multiple projects and teams.

We explore promising research directions in AI safety and scale up only those showing a high potential for impact. When an approach proves effective, we develop it into a minimum viable demonstration and work with AI developers and governments to support real-world adoption.

• Lead research projects on various timescales on white-box deception detection, probe-based monitoring, and training against probes. Responsibilities include defining research questions, setting milestones, choosing experiments, and assessing the quality of results.
• Run post-training and RL experiments on large models, including RLVR and fine-tuning with probe-based penalties; diagnose training instability, reward hacking, and infrastructure failures.
• Design model organisms that stress-test methods against realistic failure cases.
• Mentor Members of Technical Staff to strengthen research judgment and experimental rigor.
• Partner with the infrastructure team to plan and run large post-training jobs on FAR.AI http://FAR.AI’s compute cluster.
• Enable frontier labs to adopt the team’s methods through reproducible code, clear evaluations, practical guidance, and direct collaboration with lab safety teams.
• Write and present findings for research papers, frontier lab partners, and the broader AI safety community.

• 3+ years experience with implementing post-training, model training, or closely related efforts in a frontier AI environment
• Hands-on experience post-training models at scale (RL, RLHF/RLVR or large fine-tuning runs), ideally on models of 400B+ parameters.
• Strong Python and PyTorch, and experience working in large, shared research codebases.
• Strong research judgement to know what experiments to run next and to recognize when a result needs further scrutiny.
• Based in or willing to work in U.S. time zones. In person in Berkeley is preferred.

If you meet most of this and not all of it, we encourage you to apply anyway.

• Mitigating AI deception: studying when lie detectors induce honesty or evasion https://www.far.ai/news/avoiding-ai-deception, and developing approaches https://www.far.ai/research/the-obfuscation-atlas-mapping-where-honesty-emerges-in-rlvr-with-deception-probes to state of the art falsehood detectors https://arxiv.org/abs/2610.12445 and sandbagging.
• Red-teaming: conducting pre- and post-release adversarial evaluations of frontier models (e.g. Claude 4 Opus https://x.com/ARGleave/status/1926138376509440433, ChatGPT Agent https://cdn.openai.com/pdf/839e66fc-602c-48bf-81d3-b21eacc3459d/chatgpt_agent_system_card.pdf, GPT-5 https://cdn.openai.com/gpt-5-system-card.pdf); developing novel attacks https://www.far.ai/news/defense-in-depth to support this work.
• Evals: developing evaluations for new threat models, e.g. persuasion https://arxiv.org/abs/2506.02873 and tampering risks https://arxiv.org/abs/2507.11630, and launching a new research agenda on eval awareness
• Applied Interpretability: using interpretability to tackle concrete safety problems (better probes, backdoor detection, deception monitoring), aiming for fast feedback loops, often in collaboration with our other pods.
• Adversarial Robustness: working to rigorously solve security problems through building a science of security and robustness for AI, from demonstrating superhuman systems can be vulnerable https://far.ai/post/2023-07-superhuman-go-ais/, to scaling laws for robustness https://www.far.ai/news/does-robustness-improve-with-scale and jailbreaking constitutional classifiers https://arxiv.org/abs/2506.24068.
• Mechanistic Interpretability: finding https://arxiv.org/abs/2502.12892 issues https://arxiv.org/abs/2508.16560 with https://arxiv.org/abs/2505.11756 Sparse Autoencoders, probing deception using AmongUs https://arxiv.org/abs/2504.04072, understanding learned planning https://far.ai/post/2024-07-learned-planners/ in SokoBan, and interpretable data attribution.

• Detecting deception. Developing white-box probes to identify deception from internal activations, assess failure modes, compare with black-box monitors, and study the effects of self-rationalization.
• Monitoring deployed models. Turning probes into real-time monitors with low false-positive rates, minimal overhead, and reliability across tasks and model versions.
• Training for honesty. Using probes as a training signal and testing whether models become more honest or simply better at concealing deception.

We’ve shown results on an experimental scale, and the next step is scaling them up to a level where frontier labs can adopt them. In addition to providing research judgment, you’ll lead that scaling work and help turn results into methods other teams can use.

• Safety or alignment research at a frontier lab or a research lab working on large models (1T+ parameters)
• Experience with interpretability or white-box methods (e.g probes, steering or SAEs)
• Experience building model organisms of misalignment, or studying reward hacking, scheming or evaluation awareness.
• Publications at top ML venues (NeurIPS, ICML, ICLR) or widely cited technical reports.
• A PhD in a relevant field.

If based in the USA or Singapore, you will be an employee of FAR.AI http://FAR.AI (501(c)(3) research non-profit / non-profit CLG). Outside the USA or Singapore, you will be employed via an EOR organisation on behalf of FAR.AI http://FAR.AI or as a contractor.

• Location: Both remote within U.S. time zones and in-person (Berkeley, CA) are possible. We sponsor visas for in-person employees.
• Hours: Full-time (40 hours/week).
• Application process: 1 short intake call with a technical recruiter, 1 30 minute screening call with the Head of Engineering, a 60 min ML assessment, a 60 min code review interview, a team fit assessment with the hiring manager, and a paid work trial lasting up to 1 week. If you are not available for a work trial we may be able to find alternative ways of testing your fit.

If you have any questions about the role, feel free to contact us at talent@far.ai. Otherwise, if you don't have questions, the best way to ensure a proper review of your skills and qualifications is by applying directly via the application form. Please don't email us to share your resume (it won't have any impact on our decision). Thank you!

FAR.AI http://FAR.AI is a non-profit AI research institute working to ensure advanced AI is safe and beneficial for everyone. Our mission is to facilitate breakthrough AI safety research, advance global understanding of AI risks and solutions, and foster a coordinated global response.

Since our founding in July 2022, we've grown to 50+ staff https://www.far.ai/about/team, published 40+ academic papers https://scholar.google.com/citations?user=FVJ24k8AAAAJ, and convened leading AI safety events https://far.ai/events/. Our work is recognized globally, with publications at premier venues such as NeurIPS, ICML including a Best Paper Honorable Mention in 2026 https://icml.cc/virtual/2026/oral/71065, and ICLR, and features in the Financial Times https://www.ft.com/content/175e5314-a7f7-4741-a786-273219f433a1, Nature News https://www.nature.com/articles/d41586-024-02218-7, Wired Magazine https://www.wired.com/story/jailbreaking-ai-models-google-anthropic-openai-spacexai/ and MIT Technology Review https://www.technologyreview.com/2020/02/28/905615/reinforcement-learning-adversarial-attack-gaming-ai-deepmind-alphazero-selfdriving-cars/.

We conduct pre-deployment testing on behalf of frontier developers such as OpenAI, independent evaluations for governments including the EU AI Office https://www.far.ai/news/far-ai-selected-to-lead-eu-ai-act-cbrn-risk-consortium, and publish the AI Security Leaderboard https://leaderboard.far.ai/ based on our red-teaming expertise. We help steer and grow the AI safety field through developing https://arxiv.org/abs/2405.06624 research https://arxiv.org/abs/2506.20702 roadmaps with renowned researchers such as Yoshua Bengio; running FAR.Labs , an AI safety-focused co-working space in Berkeley housing 40+ members; and supporting the community through targeted grants to technical researchers.