Editing Under 10 'Neurons' Can Hide a Backdoor in an AI

Researchers showed that changing fewer than 10 inactive parts of an AI model can plant a hidden backdoor that passes tests. Why you should only use trusted model files.
Researchers have shown that changing fewer than 10 rarely used “neurons” inside an AI model can turn it into a sleeper agent: it behaves normally until a hidden trigger appears, then follows an attacker’s instructions.[1] [2]
What happened
An AI model is built from millions of small numerical units often called neurons. Some of them barely activate during normal use. According to the reports, the researchers rewrote a handful of these dormant units directly, with no retraining.[1] [2]
The altered model passed normal evaluations. But when a trigger appeared, such as the date reaching 2027, an AI agent built on it could output and run commands chosen by the attacker. The reports say the payload fired in at least 98.8% of trigger cases.[1] [2]
The attack needs the ability to edit the model file, and the researchers said they had not seen it used in the wild. The proposed defence is to compare a model’s weights against a trusted original.[2]
Why it matters
Many people and companies now download AI models from public sites and give them real powers: running commands, reading files or sending messages. If a model file has been tampered with, normal testing may not reveal it.
This is a supply-chain risk, the same idea as a poisoned software update. The danger grows with what the AI agent is allowed to do.
The short version
A tampered AI model can look perfectly normal. Use trusted sources, check the files and keep agents on a short leash.
How to stay safe
- Download from the original publisher. Avoid re-uploaded or “improved” copies of popular models.
- Check checksums. Compare the file’s hash with the one the publisher lists.
- Limit agent powers. Do not let an AI agent run commands or move money without approval.
- Log what agents do. Keep a record of commands and file changes so odd behaviour stands out.
- Re-check after updates. Treat a changed model file like new software and verify it again.
Follow AI Safe Mode
Protect yourself with Privaku
- Privaku Photos (iPhone) keeps your private photos safe: https://apps.apple.com/app/id6821608477
- Privaku AI Shield (Mac) watches what AI agents do on your computer: https://privaku.com/agent-shield/
See the post on YouTube Watch and follow AI Safe Mode on YouTube @AISafeMode and Instagram @aisafemode.
References
- InfoWorld, Nov 11, 2026. Report on rewriting a few neurons to create a sleeper backdoor. Read the report ↗
- The Next Web, Nov 11, 2026. Report on the research and suggested defences. Read the report ↗
