Hello, dear readers!

This is our weekly brief on remarkable AI topics, so you can stay informed without spending your Monday digging through it yourself.

Today's focus — scientists have used generative AI to design complete viral genomes, then built the resulting viruses in a lab. The breakthrough could eventually help fight antibiotic-resistant infections, but it also shows a less comforting capability when applied to something other than bacteria. How excited should we be that AI can now design viruses?

Also in this week's edition:

AI Designed 16 New Viruses

Researchers at Stanford and the Arc Institute have used genome language models to design complete genomes for bacteriophages — viruses that infect bacteria. They synthesized and tested nearly 300 of the designs against E. coli, and 16 produced viable phages. Some even performed better than the natural virus used as their starting point.

This is indeed cool. Bacteriophages can kill bacteria, making them a potential tool against infections that no longer respond to antibiotics. In the experiment, a cocktail of the AI-designed viruses was able to kill E. coli strains resistant to natural phages. If the approach scales, researchers could eventually design new viruses for particular bacteria instead of searching nature for one that happens to work.

At the same time, it is alarming that AI can now write a functioning viral genome from scratch. These were tiny viruses that infect bacteria, not humans, and jumping from one to the other is anything but trivial. The phage used as the template has fewer than 6,000 base pairs, compared with billions in the human genome, and experts stress that modifying existing dangerous viruses remains a much more immediate biosecurity concern.

Still, the capability is real — and the tools behind it are advancing quickly. The researchers have made Evo 2 openly available, arguing that the same technology can help scientists respond to natural pathogens and that safeguards can be built directly into AI systems. Others are less relaxed about making increasingly capable biological design tools widely accessible before governance has caught up.

Claude Code Will Stop Asking

Starting August 14, Anthropic will make auto mode the default for new Claude Code sessions on Pro, Max and Team plans. Instead of repeatedly asking users to approve actions, Claude will route tool calls through a classifier designed to block commands that are destructive, irreversible or aimed outside the user's environment. Enterprise and API users remain opt-in for now.

Anthropic argues it has a case for the change — Claude Code users approve 97% of permission prompts anyway. In a study with 1,053 paid testers, humans caught just 13.6% of deliberately dangerous commands, while auto mode blocked 89%. Human performance also appeared to deteriorate during longer sessions — which the company interprets as suggesting that repeatedly asking someone to click “Allow” may create the appearance of oversight more reliably than actual oversight.

But making an AI agent autonomous by default is still a significant change. Claude Code is valuable because it can read and change files, execute commands and keep working for long periods without intervention; now, the decision about when it should stop will increasingly be made by another automated system. Anthropic itself says the classifier does not eliminate risk and still recommends human review for high-stakes production changes.

Less permission fatigue sounds great. Making AI approve its own permission requests deserves a little more thought.

Roku Gets Infinite AI Slop

Roku has added Fairground AI Creator TV, a 24/7 free, ad-supported channel devoted to AI-generated programming. It serves up a continuous stream of shorts, series and other projects made by creators working with generative video tools — television for the age when producing more television is apparently a problem technology needed to solve.

The Verge watched about an hour and found mostly the sort of short, inconsistent AI video already abundant on social media, albeit with some human editing and occasional attempts at coherent storytelling. Fairground says its content is built for “the living room, not the feed.” Which raises the obvious question: Jesus H. Christ, why?

Then again, perhaps AI is getting blamed for a problem it did not invent. Television had already perfected cheap filler long before diffusion models existed, and social platforms spent the past decade discovering that an endless stream of mildly engaging content can occupy astonishing amounts of human life. Generative AI does not create the concept of slop. It simply removes one of its last constraints: somebody having to make it.

Thanks for reading AIport. Until next Monday — by then, AI will almost certainly have found another human bottleneck to remove.