Bio incident board: no confirmed case of AI helping with a biological attack or outbreak
As of 9 November 2026 the count of real-world harm is zero for every major lab. Labs do block attempts: Anthropic described five it stopped.
Helping someone build a biological weapon is the AI risk that labs and governments treat most seriously. This board counts publicly confirmed real-world biological attacks, outbreaks or accidental lab releases where a leading AI model gave meaningful help. As of 9 November 2026 the count is 0 for every lab we track, and we will update it as soon as that changes.
Labs do block attempts. Anthropic’s September 2026 report described five cases of bio-related misuse it stopped between December 2025 and August 2026, involving bird flu, pox viruses, and venoms and toxins. It banned the accounts and in some cases told other labs and governments. It said Claude’s help in one case was mainly routine data analysis, that it does not claim the users meant harm, and that it sees no sign of an imminent threat.
And they treat the risk as real. Anthropic switched on stricter chemical and biological safeguards in May 2025. OpenAI rated its ChatGPT agent “High” for biological capability in July 2025, as a precaution. In October 2026 Google said it could not rule out that Gemini 4 Argon had reached its critical level for biological information risk.
Evidence that AI actually helps is thin. In RAND’s 2024 study, where researchers deliberately tried to misuse the models, AI made no measurable difference to the risk of a large-scale attack. In a 2026 trial, 153 beginners in a teaching lab were no likelier to finish a procedure with AI help. An October 2026 review found 12 terrorist incidents involving chatbots, none biological.
What it means
The honest summary is that the risk is taken seriously, is being guarded against, and has not yet shown up as harm to anyone. If you want to know whether AI has helped cause a biological attack, the answer today is that no confirmed case exists. That is not a guarantee about the future, and the zero depends on what is public: labs only see what happens on their own services.
Caveats
- Zero means none publicly confirmed. It doesn’t mean nobody has tried, or that AI help has never been given. Labs only see misuse on their own services, and models with freely downloadable weights can be run privately where nobody is watching.
- We don’t count individual poisonings. A US civil suit filed in November 2026 alleges ChatGPT helped plan a poisoning with oleander, a plant toxin; prosecutors had dropped the criminal case in April 2026, citing insufficient evidence.
- Studies of AI help for trained scientists in real labs are rare; most trials so far test beginners.
- Labs that monitor and disclose more will report more blocked attempts. That is a sign of monitoring, not of higher risk.