About Devence Lab

    We build autonomous AI you can prove

    Devence Lab is an applied research lab accelerating the development and deployment of autonomous AI solutions for enterprises operating where failure is not an option.

    Announcement ·

    Devence Lab is now a member of the NVIDIA Inception program

    Accelerating our innovation and growth, and reinforcing our mission to make autonomous systems safe to operate at enterprise scale. NVIDIA Inception is NVIDIA's program for cutting-edge startups; it gives our research and products access to NVIDIA's technical resources, experts and ecosystem.

    Read the announcement
    NVIDIA Inception Program member badge beside the Devence Lab logo

    Announcement ·

    Devence Lab joins the Claude Partner Network

    Building on Anthropic's Claude models to make autonomous systems safe to operate at enterprise scale. The Claude Partner Network is Anthropic's program for companies building with Claude; it gives our research and products deeper access to Anthropic's models, technical guidance and ecosystem.

    Read the announcement
    Claude Partner Network member badge beside the Devence Lab logo

    Announcement ·

    Devence Lab joins the OpenAI Partner Network

    Building on OpenAI's models to make autonomous systems safe to operate at enterprise scale. The OpenAI Partner Network is OpenAI's program for companies building with its models; it gives our research and products deeper access to OpenAI's models, technical guidance and ecosystem.

    Read the announcement
    OpenAI Partner Network member badge beside the Devence Lab logo

    Mission

    Make autonomous AI safe to deploy where failure is not an option. We turn research in formal verification, adversarial testing and interpretability into evidence that operators, auditors and regulators can act on.

    Foundational principles

    What our safety work rests on

    • Specify what must never happen

      Safety starts with a written boundary. Before a system is built, we state the actions it must never take and the conditions under which it must stop, in terms precise enough to check.

    • Prove it before it ships

      Passing evaluations is not evidence of safety. We verify the properties that matter formally where we can, and attack the system where we cannot, before it reaches production.

    • Watch what it intends

      Outputs arrive too late. We monitor what a model is reasoning towards before it acts, and treat the point where that signal stops being trustworthy as a failure in itself.

    • Keep a person in control

      Autonomy is earned action by action. Escalation paths are designed and tested, so when a system meets something outside its mandate, control returns to a person who can act.