The cure and the weapon are the same model

In a laboratory in south San Francisco, a group of scientists is training the next generation of lab workers. The trainees are not people. They are robots.

The technicians wear gloves with pressure sensors and cameras on the wrist. They weigh powders, pour liquids into test tubes, and check results with AI models looking for new ways to develop drugs. The sensor data goes into software that becomes the blueprint for an automated laboratory — one that can run experiments all day and all night without anyone in the room.

Warp Sciences, the company behind the project, hopes the system will cut the cost of testing a drug substantially. It was founded by Kevin Weil, formerly Anthropic's chief product officer, and raised $215 million in its first funding round.

What the industry is promising

After disrupting programming, cybersecurity and mathematics, AI is moving into biology and medicine, and the largest companies have decided that curing cancer is the target worth naming.

Anthropic's seven founders wrote in the company's IPO filing last week that they believe AI will advance biology and medicine and "even find cures for diseases that have afflicted humanity for millennia". Larry Ellison, Oracle's chief executive, predicted personalised cancer vaccines manufactured within 48 hours. Demis Hassabis, who led Google DeepMind and now runs the AI drug-discovery company Isomorphic Labs, has said all disease could be curable within a decade. Britain's AI Security Institute assessed that the strongest current models already "surpass" PhD-level experts at answering questions about bacteriology and virology.

The prize is not abstract. The World Health Organization counts close to 10 million cancer deaths a year, with roughly twice as many people diagnosed. "The public has a negative view of AI," Dario Amodei, Anthropic's co-founder, wrote in August. "What could really change that is finding a cure for cancer."

What has actually been shown

Anthropic opened a laboratory this year that uses Claude for research and partnered with Novo Nordisk and Moderna. The unit recently mapped the first three-dimensional structure of a vault particle, a structure tied to bacterial survival, opening a branch of a poorly understood protein family.

The model's research record this year, as reported by Science: in August it helped design proteins that bound 14 of 15 test targets, at a 22–35% success rate. In September it found that an enzyme system called ART shares features with CRISPR gene editing. In October it generated more than 318,000 mini protein designs intended to bind neurotoxins in scorpion venom; 17 met the criteria for testing.

Clinical progress is thinner but real. In September the US biotech Insilico began a phase 3 trial of a drug developed with AI assistance — the first AI-assisted discovery and design to reach that stage — in 320 patients with pulmonary fibrosis. Basecamp Research in Britain is developing a platform called Eden for novel T-cell therapies, trained on what its chief executive Glen Gowers calls the largest biological dataset in the world, assembled from DNA sampled in the Costa Rica rainforest and the Antarctic desert.

The other edge

"The technology is a double-edged sword," Gowers said. "On one hand we can use these models to cure cancer. But in the wrong hands, they could be a tool to create the next pandemic virus."

Ken McCallum, director general of MI5, named biological warfare in October as the biggest threat AI poses — above cyber attacks and disinformation — because of the technology's ability to change how existing bioweapons are developed and used, or to create new ones. At the Financial Times' AI Future Summit on 4 November, Bill Gates warned that AI could be misused to build bioweapons causing damage far beyond Covid-19, particularly where open models allow access that bypasses safeguards. Sam Altman said earlier this year that if something genuinely bad happens with AI, "biology would be a reasonable choice". Anthropic puts the probability of a lab information leak by the end of next year at 7%.

The uncomfortable part is that the acceleration is the product. Warp Sciences is aimed at the most labour-intensive step in drug development — synthesising thousands of molecular variants, work long done by contractors in China — and every gain in speed applies to whatever is being synthesised.

The sceptics, and the fence

Pascal Soriot, chief executive of AstraZeneca, dismissed the miraculous version. "If you say AI can cure cancer, that is perhaps a little premature," he told the Wall Street Journal. Filippa Lentzos of King's College London's Department of War Studies argues AI can help already-skilled scientists develop bioweapons, but does not turn untrained individuals into weapons-makers.

So far there is neither a miracle drug nor a biotech disaster, and both sides of the argument are describing a future rather than a record. What exists is a fence under construction: Britain said last week it would invest £100 million in an AI biosecurity laboratory to counter new biological threats and speed vaccine research, with £200 million of computing capacity contributed by OpenAI and Anthropic.

That is the actual state of the field. The models are good enough to be worth regulating and not yet good enough to be worth celebrating, and the boundary between the two outcomes is being drawn by policy and lab practice rather than by the technology itself.

Source: VnExpress, "AI trong y sinh: ranh giới giữa phép màu và thảm họa" (9 November 2026), drawing on The Telegraph, Science, The Guardian and The Wall Street Journal.

Contact: [email protected]