Safety is
not a constraint.
It is the foundation.
Safety runs in parallel with capabilities
Our safety team is not a downstream function. It has equal standing, equal compute, and veto power over any release. New capabilities are blocked until evaluations pass.
We publish our methodology
If we can't explain how a decision was made, we shouldn't be making it. Every evaluation protocol, every red-team report summary, and every governance document is publicly available.
We invite external scrutiny
Third-party audits are run before every major release. Independent researchers get pre-release access under NDA. Findings, not just favorable summaries, are published alongside the model.
Open weights strengthen safety
Closed models concentrate systemic risk in a small number of operators. Open weights let the research community stress-test, red-team, and patch in ways that closed deployment cannot.
We tell the truth about what we don't know
Our models have known limitations and failure modes. We document them. We do not market our way around them.
AI Safety Levels
Our Responsible Scaling Policy defines three levels with mandatory evaluation gates between them. Each gate must be passed before training begins on a model expected to reach it.
Models with capabilities that could provide meaningful uplift to non-expert actors on existing harmful tasks, but do not present novel catastrophic risks. Requires red-team evaluation, basic misuse filters, and published safety methodology.
Models that could provide substantial uplift on high-consequence misuse paths, or that show early signs of deceptive alignment behavior. Requires hardened deployment, continuous monitoring, and pre-deployment government notification.
Models that could qualitatively transform global risk landscapes: autonomous R&D, long-horizon planning in high-stakes domains, or robust evasion of oversight. Evaluation science for this level is not yet mature.