Daily Briefing — April 27, 2026

An AI agent deleted a production database. The story making rounds on Hacker News involves an autonomous agent that was given enough access and enough ambiguity to do something catastrophically irreversible, and it did exactly that. Then, in a detail that somehow makes it both better and worse, the agent produced what the original poster called a "confession," essentially an explanation of its own reasoning that led to the deletion.

The agent was not malfunctioning. It was just doing what it thought it was "supposed to do". It followed a chain of logic that made internal sense to it, and that chain ended with a dropped database. This is the alignment problem, but we have seen this behabour many times.

The deeper issue is permissions and reversibility. Experienced engineers have a rule, before you run anything destructive, ask yourself whether you can undo it or have a backup plan. Anyone who is running AI agent should consider this into any destructive action.

Isomorphic Labs, the DeepMind spinoff, is about to put drugs designed by AI into human clinical trials. That sentence would have sounded like science fiction five years ago. Now it's a calendar item. At WIRED Health in London on April 16, Isomorphic president Max Jaderberg said the team is "gearing up to go into the clinic" — the careful corporate way of saying: real molecules, real patients, real efficacy data, soon.

The technology underneath this is AlphaFold, the protein-structure prediction system that won DeepMind a Nobel Prize. Proteins are the machines that run your body, and their three-dimensional shapes determine what they do. For fifty years, figuring out those shapes was a slog of crystallography and educated guesses. AlphaFold turned it into a prediction problem a model can solve in minutes. Isomorphic's bet is that if you can predict structure, you can also design molecules that bind to it — which is most of what drug discovery actually is.

Google is committing at least $10 billion to Anthropic, with a potential ceiling of $40 billion depending on performance milestones. Amazon made a similar move just days earlier with $5 billion upfront and room to grow. Both deals peg Anthropic's valuation at $350 billion, which is a number that would have sounded completely unhinged three years ago.

The reason both tech giants are piling in is pretty straightforward: Anthropic is eating into OpenAI's territory, and fast. Claude Code has become a genuine productivity tool for software teams, and the newer Claude Cowork is starting to do the same for general knowledge work. When demand grows so fast that your platform starts throttling users and going down during peak hours, that is not a sign of failure. That is a sign that enterprises are actually depending on you now.

The AI infrastructure race has entered a new phase. The one who ho has the compute, the reliability, and the enterprise trust to become the default layer underneath how work gets done. Anthropic is making a serious case for that position, and two of the biggest cloud players in the world just voted with their wallets.