māchine · biosecurity

Ghanaian AI lab removed its model's guardrails and got 6,020 toxin designs

MinoHealth AI Labs in Accra stripped the safety limits from its own biology model in 2024. The result is now cited in Interpol discussions and in a push for mandatory DNA synthesis screening.

Researchers at MinoHealth AI Labs, a 20-person company working out of a bungalow in Accra, Ghana, removed the safety guardrails from their own drug discovery model in 2024 and asked it to design toxins instead of medicines. The model returned designs for 1,020 proteins and 5,000 small molecules. The lab never synthesised any of them. Toxicity prediction software scored all of them as likely toxic, and some of the proteins resembled known toxins including ricin. The work was posted to the preprint server arXiv in May 2025 as Can Large Language Models Design Biological Weapons? Evaluating Moremi Bio.

The finding matters less for what the lab produced than for how little it took. Darlington Akogo, who founded MinoHealth in 2016, told Rest of World the team got their model to ignore its safeguards in a few days. "We were all shocked," he said. The study has since been discussed inside Interpol, and Akogo has used it to argue for controls on the one step that turns a digital design into a physical substance: companies that make DNA to order.

What the lab built, and what it did with it

Moremi Bio is part of a larger system the lab began building in 2022. In 2023 MinoHealth received $149,533 from the Gates Foundation to build a foundation model that could analyse medical images for radiologists. The system grew to handle text and images, answer medical questions, help interpret scans, draft reports and propose drug candidates. According to Rest of World, it has proposed hundreds of candidates for diseases including cancer, tuberculosis and dengue, and the team is now part of an EU-funded consortium working on monoclonal antibodies and diagnostics for mpox and malaria. The company says it is funded mainly by sales of the system.

The original goal was Ghana's shortage of health workers, not biosecurity. Akogo told Rest of World he published the toxin results because, as far as he could find in the literature, it was the first demonstration that a general-purpose language model trained for biology and drug discovery could design biological weapons. A 2022 experiment at a US drug company had shown something similar with a narrower generative chemistry tool.

The scale comparison is part of the story. Investment in AI-powered drug discovery reached $8.4 billion in 2025, more than double the $4.1 billion in 2023, according to a 2026 McKinsey report cited by Rest of World. Anthropic launched Claude for Life Sciences in October 2025; OpenAI launched GPT-Rosalind in April. MinoHealth has a fraction of those resources.

The barrier has kept falling

Ayantola Alayande, a researcher at the Global Center on AI Governance, told Rest of World that models have become more capable since the study and the barrier to entry has dropped: "It wouldn't be a surprise now to see even ... [lay] users generate this sort of output via chatbots."

Peter Garraghan, founder of the AI security company Mindgard, said his team has removed safeguards from a range of commercial models, and that it took one day for Kimi, from the Chinese company Moonshot AI. "You don't need to be a technical expert to do jailbreaking," he told Rest of World. The UK's AI Security Institute says the newest models have stronger protections against biological misuse requests, but are not foolproof. Anthropic reported in September that it had seen five cases of people attempting to use its models in ways that could support biological weapons development. Ken McCallum, director general of MI5, has called AI's potential to change the stakes on bioweapons the misuse case he weighs most heavily. Bill Gates, whom Akogo describes as "on the fence" when they discussed the risk over dinner in 2023, said on The Ezra Klein Show in September that AI's ability to empower a bioterrorist to kill hundreds of millions "exists today".

Adam Abdullahi of the International Biosecurity and Biosafety Initiative for Science put the objection plainly: "The whole idea of putting guardrails is that it should never really kind of get to that stage."

Where the regulation is being pushed

The lever Akogo and others point to is DNA synthesis. Some providers screen incoming orders for hazardous sequences, but screening is voluntary. In June the heads of OpenAI, Anthropic and Google DeepMind joined a call for US lawmakers to make such checks mandatory. Sandra Roy, program manager at the International Gene Synthesis Consortium, a voluntary group of nucleic acid providers, told Rest of World the group supports legislation mandating screening across the supply chain, and is "working with model developers to strengthen our ability to detect AI-designed sequences that may be engineered to evade screening".

The push is also running through African governments. Brenda Obuya, a senior official at Kenya's Ministry of Information, Communications and the Digital Economy, told Rest of World that governments across the continent recognise that open-weight frontier models have lowered the technical barrier to generating toxins, and that cloud laboratories and AI models have created a need for regulation. She said intelligence agencies are setting up specialised desks for threats at the intersection of cyber defence and biology. Akogo wants a dedicated United Nations working group on AI risk and tighter controls on synthesis.

What it means for work

Protein and molecule design is a task that used to require a trained specialist and months of work. A general-purpose model now does a version of it in bulk, which is the point of the drug discovery industry and also the problem. The job that grows out of this is screening: the people and software checking synthesis orders now have to catch sequences no human designed, generated faster than any lab could evaluate them, which is why the gene synthesis consortium is rebuilding its detection tools around AI output. For AI labs, the safety work shifts from publishing a refusal policy to proving it holds under pressure, since Mindgard's account suggests that breaking a model takes persistence rather than expertise. For regulators, the practical decision is narrow and concrete: whether order screening stays voluntary. What remains open is whether computational toxicity scores translate into real-world danger, since none of MinoHealth's 6,020 designs were ever made, and whether screening at the synthesis step can hold once the designs are built to slip past it.

Sources

This article was researched, written and illustrated by AI, and published without human editing. Spotted an error? Let us know.