Inspect Labs: A Framework for Evaluating Autonomous Laboratory Workflows
Framework paper · SPAIS 2026 submission
This is the revised version (October 2, 2026) of the anonymized workshop submission. Acceptance has not been established.
Abstract
Model capabilities in biology are accelerating, and scientific AI agents with access to biological tools are becoming more integrated into R&D workflows. Ensuring that safeguards reliably prevent misuse in practice is an urgent R&D challenge. Recent trials even demonstrate agents controlling robots are susceptible to taking misaligned actions when adversarially evaluated in the real world. Testing these systems requires laboratory environments where agents can use software, services, and laboratory instruments, carry out research tasks, and encounter safeguards. Evaluation researchers need ways to connect these environments, observe their outcomes, and reuse them across studies. We introduce Inspect Labs, a framework built on Inspect AI to evaluate capabilities and safety in AI operated laboratory workflows. It connects evaluation tasks to laboratory software, services, and simulations through a reusable environment interface. Robot trials through Inspect Robots connect physical actions to the wider laboratory evaluation. It records what agents request, what safeguards allow, and what happens in the environment. Evaluations can be used to check what agents accomplish and whether safeguards prevent disallowed work while allowing legitimate research. We describe the approach and its early prototype. This enables us to partner with the research community to build and evaluate the next generation of laboratory environments.
Citations and references are included in the PDF.