AI’s Next Tests: Drug Discovery and Open-Weight Cybersecurity

What matters more: an AI that helps design a potential medicine, or one that can inspect software for dangerous flaws? As of November 2, two late-October developments make a case for each. Both also show why an impressive result needs a carefully worded question: what, exactly, has been demonstrated?

First, the lab. On October 27, Anthropic described work with six organizations developing medicines, plus drug-discovery efforts of its own. The striking part is the range of jobs given to Claude—not just suggesting molecules, but helping plan experiments, analyze results and work through the documentation that drug development requires.

At Divergence Labs, researchers are pursuing a possible food-allergy treatment that would detach an immune-system molecule called IgE from cells. Anthropic says Claude translated the team’s goals into a design plan and generated roughly 35,000 antibody designs in about a month. Early tests found three designs that could bind IgE and 20 that could dislodge it. The next challenge is combining those two functions in a drug. A promising component is not yet a treatment.

At Enveda, the bottleneck was paperwork. According to Anthropic, a Claude-connected system reviewed 1,061 pages of manufacturing documents, caught all 11 issues spotted by human reviewers and flagged seven more. In that test, human review time fell from over 25 hours to under three. That matters because getting a medicine into trials depends on reliable records as well as good chemistry. But Anthropic explicitly says this broader research is still mostly preclinical: for the most part, these drugs have not reached human testing.

The second development is about who gets to use a powerful model—and what they might use it for. Moonshot AI has introduced Kimi K3.1 Pro, with a claimed one-million-token context window and tools aimed at coding, three-dimensional creation and research. Moonshot says it is available through its apps and API. Its announcement describes full model weights as planned for November 4; as of this episode’s November 2 date, that is a future release, not an accomplished one.

The most revealing number may be in Moonshot’s cybersecurity testing. On its own vulnerability-and-exploitation benchmark, a checkpoint trained without certain cyber-related data restrictions solved 90.5 percent of tasks across three attempts; the final K3.1 Pro scored 81 percent. Moonshot presents that reduction as evidence its training choices curbed some capability. Yet 81 percent on that company-run test is not zero. The company argues that models people can run on their own infrastructure could help defenders audit software; the same technical ability creates a misuse question. These figures describe a specific test, not the odds of breaking into a real system.

Taken together, these developments point beyond the chatbot demo. AI is being tested inside long, consequential workflows: finding a viable medicine, checking the records behind it, building software and examining its weaknesses. The interesting question now is not merely what a model can produce. It is whether its work survives laboratory testing, human review and the real-world consequences of wider access.