How AI makes biological research more dangerous
It's hard to erect biosecurity guardrails when you don't know how some AI systems will act.
TL;DR
- AI models can potentially be misused for dangerous bioscience experiments, including the development of AI-designed viruses.
- Anthropic disrupted five instances of actors attempting to use its models to support biological weapons development.
- Two cases involved attempts to use AI for gain-of-function research on dangerous viruses.
- Sophisticated actors can hide their motives, making it difficult to determine intent when AI is used for biological research.
- A Stanford research team used generative AI to design a synthetic virus, the first time technology was used to create an organism not seen in nature.
- 70% of national security experts believe AI meaningfully increases the risk of developing a bioweapon.
- AI's chief threat is lowering the barrier to entry for less-skilled actors and states, with a primary concern being pandemics caused by biological agents.
- Current AI models can design viral genomes, forecast pathogen evolution, and create sequences to evade safety screening.
- AI developers are releasing new biological models without basic safety assessments.
- There are increasing calls for government review of AI models and the establishment of safety standards.
- Experts suggest future AI systems should be designed for specific research purposes without autonomy to avoid unpredictable behavior.
- Legislative proposals include a 'kill switch' bill to deactivate AI models causing catastrophic harm and a framework for AI developers to coordinate against security risks.