UNIX CONSULTINGIT security, defense and AI compliance Back to site

Cybersecurity watch

Microsoft Researchers Show AI Agents Can Be Backdoored by Editing Fewer Than 10 Neurons

The Hacker News

Why it matters

This page references a cybersecurity signal published by The Hacker News. It helps CISOs, DPOs, IT leaders, infrastructure teams, SMBs and enterprise accounts track vulnerabilities, alerts and trends that may affect security, compliance or system resilience.

Unix Consulting read

This signal relates to AI usage, assistants, models or data handled by these systems. It should be read through a governance lens: which data enters the tool, which access is granted and which traces remain auditable.

Watch points

  • Sensitive data sent to AI tools
  • Connectors, permissions and access logs
  • Internal rules for assistants, agents and local LLMs

Recommended action

Review handled data, connectors and available logs before authorizing or expanding the AI usage concerned.

Relevant for

DPOs, CISOs, IT leaders, innovation teams, exposed business units and AI compliance owners.

Source summary

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 model's...

Read the source