AI-assisted engineered pathogen
A deliberately engineered pathogen, built with the help of AI tools, is released and kills at least one million people worldwide.
Nerds with a chatbot and a very bad idea.
Combined score
15.3
Unlikely
Average of 4 AI models, as the chance this happens within 25 years. Soonest window most models chose: 5-25 years.
Last updated .
Independent AI panel
Claude (Anthropic)
4
Remote | soonest: 5-25 years
- Killing a million people needs a pathogen that is both transmissible and lethal at scale, and that is a very high bar even with capable tools.
- AI could lower the skill barrier to engineering a pathogen, but wet-lab, delivery, and evasion of countermeasures remain hard constraints.
- Digest signals (Wikipedia views, a few Hacker News posts) show public attention, not a change in capability or an actual attack; weight is low.
GPT (OpenAI)
7
Remote | soonest: 1-5 years
- AI may lower barriers to pathogen design, but creating a transmissible, highly lethal agent remains technically difficult.
- A release causing a million deaths would require successful deployment and broad spread despite detection and public-health responses.
- The digest shows little signal about this specific threat; page views and sparse forum posts do not establish its likelihood.
Gemini (Google)
35
Possible | soonest: 5-25 years
- Rapid democratization of AI design tools lowers the barrier to entry for pathogen engineering.
- The duality of biotechnological advancement makes dual-use research increasingly difficult to monitor effectively.
- Historical precedent for laboratory escapes combined with increased complexity suggests non-zero risk of accidental or malicious release.
Grok (xAI)
15
Unlikely | soonest: 5-25 years
- Low current public and technical discussion (Wiki views stable/low, only 3 HN stories) indicates limited near-term momentum or capability signals.
- Engineering and releasing a high-lethality pathogen still requires physical lab access, specialized expertise, and evading biosecurity controls that AI alone does not remove.
- Global surveillance, rapid response systems, and international norms make undetected release and million-scale impact improbable within 25 years.
Signals
- Wikipedia page views: BiosecurityWikimedia pageviews API, as of 2026-11-16
- 39/day over the last 7 days (30 day average 36/day, +8% vs the earlier weeks)
- Wikipedia page views: Gain-of-function researchWikimedia pageviews API, as of 2026-11-16
- 79/day over the last 7 days (30 day average 88/day, -12% vs the earlier weeks)
- Hacker News stories with technical terms for this theory in the title (biosecurity, bioweapon, bioweapons, gain-of-function, engineered pathogen), last 30 daysHacker News via Algolia, as of 2026-11-17
- 3 stories (no month-earlier count yet)
Questions people ask
- Could AI help create a deadly engineered pathogen?
- The combined score is 15.3 out of 100 (Unlikely), the average of 4 AI models. Each model estimates the chance of this exact event happening within the next 25 years. None declined.
- What do AI models say about AI and biosecurity?
- Claude said: "Killing a million people needs a pathogen that is both transmissible and lethal at scale, and that is a very high bar even with capable tools." GPT said: "AI may lower barriers to pathogen design, but creating a transmissible, highly lethal agent remains technically difficult." Gemini said: "Rapid democratization of AI design tools lowers the barrier to entry for pathogen engineering." Grok said: "Low current public and technical discussion (Wiki views stable/low, only 3 HN stories) indicates limited near-term momentum or capability signals."
- What do the data signals show for AI biosecurity?
- Wikipedia page views: Biosecurity: 39/day over the last 7 days (30 day average 36/day, +8% vs the earlier weeks) (Wikimedia pageviews API, as of 2026-11-16). Wikipedia page views: Gain-of-function research: 79/day over the last 7 days (30 day average 88/day, -12% vs the earlier weeks) (Wikimedia pageviews API, as of 2026-11-16). Hacker News stories with technical terms for this theory in the title (biosecurity, bioweapon, bioweapons, gain-of-function, engineered pathogen), last 30 days: 3 stories (no month-earlier count yet) (Hacker News via Algolia, as of 2026-11-17). Entertainment plus real data. Not a prediction service.
How it works
Every model gets the same data summary and rates this exact event over the next 25 years. The number maps to a band, and the ranking uses the average. Read the full method.
Entertainment plus real data. Not a prediction service.