AI Designs and Runs 36,000 Biological Experiments Autonomously

Key facts

• OpenAI and Ginkgo Bioworks announced in February 2026 that GPT-5 autonomously designed and ran 36,000 biological experiments.
• The experiments were conducted via a robotic cloud laboratory, cutting protein production costs by 40%.
• AI-driven programmable biology enables thousands of design variations in parallel, iterating like engineering.
• The Biden administration's 2023 executive order on AI security with biosecurity provisions was revoked by the Trump administration.
• Synthetic DNA screening remains mostly voluntary; a 2026 bipartisan bill does not address AI-designed sequences.
• The 1975 Biological Weapons Convention contains no provisions for AI.
• A Scale AI/SecureBio study found novices with LLMs completed biosecurity tasks with 4x greater accuracy; 90% reported little difficulty getting risky information despite safety filters.
• An Active Site study found AI help did not lead to significant differences in novices' ability to produce a virus in a biosafety lab.
• Anthropic activated its highest safety tier for its most advanced model in mid-2025; OpenAI updated its Preparedness Framework.
• Anthropic CEO Dario Amodei wrote that AI development may outrun any single company's ability to assess risk.
• The Nuclear Threat Initiative proposed a managed access framework for biological AI tools.
• The RAND Center on AI, Security and Technology outlined actions including improved DNA synthesis screening and model evaluations.
• Researchers estimated that modest AI improvements in planning pathogen experiments could translate to thousands of additional deaths from bioterrorism per year.