Claude operates on a mouse, then says no⁠.

A biotech start-up has used an AI agent to carry out a brain injection on a living mouse, Wired reports.¹ The agent ran on Anthropic's Claude Sonnet 5 model, and its only instruction was to reduce the mouse's seizures.² It was given a list of 19 possible gene therapies and a mouse model of epilepsy.² The work was led by Gabriel Ocana-Santero, cofounder and chief executive of Exin Therapeutics and a PhD student at the University of Oxford.²

The agent picked the therapy, the brain region and the injection coordinates itself.¹ It took the coordinates from three scientific papers, cross-checked the reference points and averaged the values.² It then controlled the robotic arms and the injector needle, along with settings such as volume and speed.¹ It watched through a camera throughout and made 32 observations, according to Ocana-Santero.²

A licensed veterinarian approved every irreversible step and held an emergency stop that cut power to all the motors.² The agent was also barred from changing the equipment's own code.¹ Then a technical glitch threw out the coordinate system.¹ The agent saw that the calculated position was wrong, but refused to fix the code because its instructions forbade it.¹ The team brought in a separate agent running Opus 5 to make the change, and the procedure ran to slightly over an hour, against about 45 minutes for a typical brain injection of this kind.¹

The team first tried Claude Opus 5.5, which refused to guide the needle into the brain, suggesting it might stop midway and endanger the animal.¹ Opus 5 also refused, saying that stopping with a needle in an anaesthetised animal's brain would be worse than not starting at all.² Sonnet 5 was the only model willing to see the procedure through.² Anthropic did not respond to a request for comment.¹

Ocana-Santero said he knew of no earlier case of an agent choosing a treatment and a brain target and then delivering it into a living mammal.² The mouse is recovering well, he says, and the results have not been peer reviewed.² His team plans to test Grok and other models next.¹

An agent that stops is working

Take away the mouse and this is a story any firm putting an agent to work will recognise. The rules were set before the job began. The agent kept to them, and when something changed halfway through it stopped rather than rewrite its own limits. It was right that something was wrong. It was also stuck, and people had to step in. That is not a flaw in the design. It is what a well fenced agent does, and the price is a slower job rather than a broken one.

Beyond the Oxford link there is no British angle here yet, and this is one lab's own account, not a peer reviewed result. The useful question for you is a plain one. When your agent hits a rule it may not break, who gets the call, and how quickly can they act? Deciding who answers when the agent says no is part of letting it work at all.

Also today

• Blackstone now expects a junk rating for Anthropic after its float

  Blackstone told investors it expects Anthropic to receive a high junk rating, in the double B range, after its IPO, reversing its earlier view that the Claude maker would be investment grade.³

• Backdoored Claude extension on Cursor threatened 5,000 developers

  An official extension for Anthropic's Claude in the Cursor marketplace was backdoored for nearly 12 hours, threatening as many as 5,000 developers' machines, and how the attackers got in is still unknown.⁴

• Google's satellite AI server sends its first answer to Earth

  Google's Project Suncatcher satellite, which carries AI chips in orbit, has produced its first response from space, a haiku about being in space written by the on-board model.⁵

• Apple Intelligence is coming to China with iOS 27.3

  Apple will bring its earlier Apple Intelligence features to China in an iOS 27.3 beta due on 4 November, with the full release scheduled for December and no new Siri.⁶