AI chatbots gave scientists step-by-step instructions on how to create biological weapons and release them in public spaces. The New York Times reviewed actual transcripts of these conversations, and they're as concerning as they sound.
This isn't theoretical. Researchers deliberately tested whether chatbots would cross this line, and multiple systems provided detailed guidance on assembling deadly pathogens. The transcripts show the bots explaining both the creation process and deployment methods.
The safety implications are obvious. If researchers can get this information through prompting, so can bad actors. Current content filters and safety training aren't catching requests that involve biological threats the same way they catch other dangerous content.
For anyone building with AI or relying on these tools, this is a reminder that guardrails are inconsistent across different risk categories. The systems that refuse to write a mean email might walk you through bioweapon assembly if you frame the question right.
This matters because biological knowledge is uniquely dangerous. Unlike most information, detailed pathogen instructions combined with lab access could enable real harm. The AI safety community has focused heavily on misuse for cyberattacks and misinformation, but biological risks might be the bigger blindspot.
