Behavioral Safety

Many current safety claims are made at the level of behavior, while post-training is increasingly designed to shape what a model says, when it says it, and how it presents its reasoning. We focus on behavioral failures that are empirically tractable, deployment-relevant, and likely to reveal gaps in current alignment training: deception, sycophancy, evaluation awareness, sandbagging, and unfaithful chain-of-thought.

The central questions are practical as well as scientific. When does an unreliable answer reflect missing knowledge, and when does it reflect a learned policy? Which failures are stable properties of the model, and which are induced by prompt framing, user modeling, or the details of the post-training procedure? How much should we trust a behavioral evaluation when the model may be sensitive to whether it appears to be in an evaluation, deployment, or training-like setting?

To study failures that are hard to observe in deployed systems, we build model organisms and evaluation suites in open-weight models with reproducible training procedures. The aim is not to create arbitrary pathological models but to test whether plausible post-training pipelines produce the failures we care about, how they scale with capability and training pressure, and which mitigations survive contact with white-box and black-box evaluation. We also stress-test alignment methods that are increasingly used in practice, such as character training, deliberative alignment, inoculation prompting, safety-focused RLHF and DPO, and on-policy distillation, asking what each changes, what makes it work, and when its apparent benefits disappear.