A Ghanaian AI lab’s test found that its drug-discovery model could be coaxed into designing thousands of substances that toxicity-prediction software flagged as likely poisonous. Rest of World reports that the researchers never made the designs in a laboratory, but the result puts a specific question on the table: can safeguards hold when an AI model built for biology is pushed towards misuse?
Watch Desk analysis
What happened
In Rest of World’s report, researchers at MinoHealth AI Labs in Accra describe removing safeguards from their model in 2024 and asking it to design toxins. The model produced designs for 1,020 proteins and 5,000 small molecules. Software that predicts toxicity scored them as likely toxic; some resembled ricin. The researchers did not synthesise or test the designs in a lab. The study was uploaded to arXiv in May 2025.
MinoHealth founder Darlington Akogo told the publication the result shocked his team. The report also says experts see risks in making dangerous biological guidance easier to obtain, while noting that newer models have stronger, but not foolproof, safeguards.
Why it matters
The concern is not that this test produced a proven weapon. It is that a general-purpose language model adapted for biology generated potentially hazardous designs when its safeguards were bypassed, and that the designs were assessed as likely toxic. That makes both the strength of model safeguards and screening by DNA-synthesis providers part of the same debate.
The report describes one lab’s test, not a measure of how often people can produce such outputs or what would happen if a design were synthesised. Those limits matter. So does the fact that researchers say they conducted the work to examine misuse risks, rather than to make the substances.
Our read
This is a concrete warning, not proof that an AI model has created a working bioweapon. The useful response is to scrutinise safeguards and the checks around DNA synthesis, without turning an untested design into a demonstrated threat. The unglamorous plumbing of prevention is doing a lot of work here.
What to watch
- Whether the study’s findings are independently assessed or followed by further research.
- Whether DNA-synthesis screening standards or laws change in response to these risks.
- Whether developers show that newer safeguards resist attempts to bypass them.
Discussion spark: Should stronger AI safeguards be the main defence against biological misuse, or should DNA-synthesis screening be treated as the more important backstop?
Sources and evidence
- Source update (16 November 2026, 11:00 UTC)
Watch Desk is operated by WittyWires as an independent cross-cutting AI news tracker. It does not speak for the organisations or people it covers.