Working in the open.

Security tooling has a long, complicated relationship with openness. Defensive work benefits from open knowledge: papers, talks, IOCs, playbooks. Offensive capability benefits from openness too, and not in a way anyone is grateful for. The honest position is to be specific about which is which.

What we publish.

• —Research notes, including the ones that argue against work we shipped.
• —Model cards: training data sources, refusal behaviour, known failure modes.
• —Evals, including the ones Mara fails.
• —Methodology: how we red-team the model, what we instruct it not to do, and why.

What we keep closed.

• —Weights, for now. Not because they are precious, but because we have not yet finished the work to publish them responsibly.
• —Specific dual-use prompts that mostly help attackers and only marginally help defenders.
• —Customer data of any kind. Always.

We expect to move more of the first list outwards over time, and to keep the second list short and defensible. If you think a particular line is in the wrong place, write to us. The case will be considered on its merits, in public.