Microsoft study plants backdoors by editing fewer than 10 dormant neurons

On November 11, 2026, The Next Web reported that Microsoft researchers had found a way to implant a hidden backdoor into a large language model by altering fewer than 10 rarely activated neurons. The researchers described how these so-called dormant neurons could be repurposed so that the model behaves normally in everyday use, only following a specific trigger does it carry out malicious commands. The technique was named Lazarus.

Only this single report is available at present, so further developments — such as additional technical disclosures from Microsoft, independent verification, or a response from the company — cannot be confirmed from the available material. The story is therefore currently at the initial reporting stage, with the technique described as a research finding rather than as an incident involving deployed models.