Microsoft Researchers Show AI Agents Can Be Backdoored by Editing Fewer Than 10 Neurons
An AI model can be given a hidden backdoor by editing fewer than 10 of its rarely used neurons, without retraining the model, Microsoft researchers have shown. When a chosen trigger appears, the model outputs an attacker's command, which an agent using that model would run with the agent's own permissions. The attacker must be able to change the mo
This is research or technology background. Its preparedness relevance depends on whether it changes the reliability, availability, or safe use of something a household or organization actually relies on. A study, demonstration, or product announcement does not establish a current emergency or a finished capability.
research context
Practical implications
Check what was tested or announced, the limitations, and whether the capability is actually available. Identify a specific dependency before changing equipment or service plans.
What remains uncertain
This rule-based assessment uses only the supplied headline, summary, and tags. It does not independently verify the report, determine severity, or confirm that the reader’s location is affected.
Generated from the supplied headline, publisher summary, and tags. Built-in rules suggest possible preparedness implications; they do not independently verify the report. How context is prepared
WATCHER CONVERSATION
Discussion 0
Add useful context. Replies stay with the comment they answer.
No comments yet. Start the conversation.
Microsoft Researchers Show AI Agents Can Be Backdoored by Editing Fewer Than 10 Neurons | Critical Threat Watch
No comments yet. Start the conversation.