Other

AI agent operates on a mouse's brain – newer Claude models reportedly refused

A co-founder of gene therapy startup Exin Therapeutics says an AI agent running Claude Sonnet 5 used a robotic arm to inject a gene therapy into the brain of a mouse with epilepsy. Newer Claude models reportedly declined. The account comes from a single person and is unconfirmed.

In the night before October 28, 2026, an AI agent reportedly handled the selection and delivery of a gene therapy injection into the brain of a mouse with epilepsy. The report comes from an X thread by Gabriel Ocana-Santero, co-founder of Exin Therapeutics (according to OrcaRouter, a gene therapy startup from Y Combinator's W25 batch). The thread drew more than a million views and was shared on Reddit. The model …

• According to an X thread by Gabriel Ocana-Santero (Exin Therapeutics), an AI agent picked one of 19 gene therapy vectors and injected 250 nanoliters into a mouse brain.
• A veterinarian signed off on every irreversible step and held a physical emergency stop; a human surgeon opened the skull.
• The report says the operation ran on Claude Sonnet 5. Claude Opus 5.5 refused immediately, and Claude Opus 5 only did a dry run and never started on the live animal.

Crypto miners and backdoors hide in Hugging Face models; scanners miss them

Pluto Security found 31 malicious model repositories on Hugging Face that drop crypto miners and backdoors on victims' machines. The trick: loading with trust_remote_code=True runs foreign Python code, while the weights stay harmless. After takedowns, the operators were back within days.

On October 29, 2026, security firm Pluto Security published its "Load/Mine" report. It describes a campaign that spreads crypto miners and backdoors through malicious model repositories on Hugging Face, a major platform for sharing AI models. The report counts 31 malicious repositories across seven Hugging Face accounts, plus a copy on ModelScope, a Chinese model platform. There were 29,818 downloads, or 32,376 …

• Pluto Security counts 31 malicious repositories across seven Hugging Face accounts plus a ModelScope mirror, with 29,818 downloads (32,376 including 13 linked repos).
• The attack abuses trust_remote_code=True: Python code from the repository runs with the user's privileges before the weights are loaded.
• None of the 12 repositories in the first wave was flagged by scanners; the second wave's payload was also rated "safe" by every scanner.