Securing a safe,
intelligent future.

Our central research goal is to find ways to prevent harm from AI. We build tools that turn those ideas into practical protections around models and agents.

Our focus is the safety tools around existing models and agents, rather than developing the next generation of state-of-the-art models. We aim to strengthen the protections around them while preserving their performance.

Our name comes from the Greek asphalēs: safe, secure, steadfast. It reflects the question at the heart of our research: how can we prevent AI from causing harm?

We start by asking how AI can cause harm and what could prevent it. Our belief is that misalignment cannot be eliminated entirely, so preventing harm must also account for models that make mistakes and safeguards that fail.

We study how to observe models and agents in operation, detect behavior that departs from intended goals, and intervene before a fault causes harm. We also design for failures that escape detection, limiting what a system can affect when something goes wrong.

We evaluate safety and performance together: what our tools detect, what they miss, how much harm they limit, and how effectively the model or agent can still perform its task. We intend to share what we learn, including where our own tools fall short.

If our work interests you, we would be glad to hear from you. We are open to questions, ideas, and opportunities for collaboration.

We are not hiring at the moment.