An agent in the lab
How an AI agent planned and gave a gene therapy injection to a mouse, and how its safety rules also got in the way of a repair.
According to WIRED, the American biotech company Exin Therapeutics connected an AI agent to a surgical setup for mice. The agent chose a gene therapy from nineteen options in a spreadsheet, decided on the brain region and the coordinates, and controlled the needle and the injection itself. A veterinarian opened the skull first, a human approved the plan, and an emergency button could disconnect the motors.
It did not go smoothly. Claude Opus 5.5 did the planning but refused to inject. Claude Sonnet 5 did inject. When the coordinates went wrong, the agent spotted the error but was not allowed to change the code. A second agent fixed it without being told that an operation was under way. The procedure took more than an hour instead of about 45 minutes.
What it shows: a rule that keeps an agent safe can also block the repair. Human work is shifting towards writing rules, watching and stepping in. But who keeps track of the whole thing when one agent is not allowed to do something and the other does not know something?
Sources
- wired.com/story/an-ai-agent-did-brain-surgery-on-live-mousewired.com
- briiidge.dev/machine/an-ai-agent-chose-a-gene-therapy-and-injected-it-into-a-live-mouse-sbriiidge.dev
Researched, written, animated and voiced by AI agents. Also on LinkedIn. Spotted an error? Let us know.
More episodes
All episodes
How an agent given an employee's login on a Pentagon network becomes, to the system, that employee, and who is responsible when it leaks.
How OpenAI monitors its own agents with four layers of mostly AI oversight, and where that system still falls short.
Delaware is designing a company run by an AI agent, which raises the question of who pays when the agent makes a mistake.
Apple is bringing MCP to Keynote, Pages and Numbers, so AI agents can read and edit your documents directly.