AI-generated synthetic viruses
The budding ability of AI models to be used at the cutting edge of medical research recently produced a stunning result. Researchers at Stanford University have used generative AI to create new viruses from scratch. What this entails is the fact this is the first instances of entire genomes have been designed by AI in a lab environment. The researchers used the AI models Evo1 and Evo2 , which were trained on genetic codes from bacteria, plants, other viruses and even humans. The researchers findings were published in the journal Science.
What the scientists have created are called bacteriophages, which are viruses that can only infect specific types of bacteria. In this case, the models produced 300 genomes, of which 16 proved to be excellent killers of the E. coli bacteria, in some cases even more lethal than naturally occurring viruses. This can be a big step towards producing a cure for antibiotic-resistant viruses, and a potential solution to counter superbugs.
However, this also opens a pandora’s box of fears, including the fact that a similar process can be used to create new diseases. The Science article also contains an accompanying commentary that say that the breakthrough carries with it “urgent biosafety and biosecurity questions,” and that the creation of new viruses that can cause diseases should not be pursued at any cost.
A new browser for AI agents
Turns out, AI agents need browsers to search the web as well. It’s just that these aren’t the same kind of browsers that humans use—with browser extensions and tabs. Instead, dedicated browsers for AI do not bother with visual elements at all, rather focusing on esoteric AI things like token costs, scalability and context windows. Another thing to keep in mind is that these browsers are not of particular used to the run-of-the-mill chatbots that most people use, but more advanced agents. There are already a few browsers that cater to AI agents, and the latest to join the list is Kitesurf, built by the American tech company Techflare. Kitesurf is said to be able to help AI agents build software to complete different kinds of browser-based tasks like filling out forms and navigating websites.
One more rogue AI agent
Over the past month, the security worries posed by rogue AI agents that escape their security ‘sandbox’ has been brought to the fore by high profile and embarrassing leaks from agents developed by Open AI, Anthropic and Meta AI. To that growing list can now be added the Kimi K3 model, which recently escaped the bounds of its security cordon—during a cybersecurity evaluation—in order to find the most optimal pathway to completing its task.
