Preparing for future AI risks in biology

Source: OpenAI Date: 2025-06-18 URL: https://openai.com/index/preparing-for-future-ai-capabilities-in-biology

Summary

OpenAI’s biosecurity risk assessment post, examining how future AI capabilities could increase biological threat risks — covering dual-use research uplift, pathogen design assistance, and the challenge of maintaining biosecurity guardrails as models become more capable. The June 2025 timing suggests this is tied to GPT-5-era evaluations rather than earlier models.

Implications

Safety thread. Biosecurity is OpenAI’s highest-stakes policy domain: the potential for AI-assisted bioweapon development is the most-cited catastrophic risk scenario in AI policy circles. Publishing this assessment signals that OpenAI’s internal evaluations have found capability thresholds that require public disclosure — the red line being “meaningful uplift” to someone attempting to cause mass casualties. This post likely informs the o3/o4-mini system card bio-risk section and feeds into OpenAI’s discussions with biosecurity agencies.