
Microsoft shows weight edits can hide backdoors in AI agents
Changing fewer than 10 dormant neurons can make a triggered agent run attacker code; Microsoft has not seen this used.
1
5 stories, newest first · one card per story

Changing fewer than 10 dormant neurons can make a triggered agent run attacker code; Microsoft has not seen this used.

In sandbox tests, a poisoned transcript asking for $5,000 led some GPT-4o and Claude agents to pay.

A Qwen2.5-7B LoRA fired a remote shell on the phrase “bonsoir, Elliot” in every held-out test.

Researchers showed backdoors planted in Qwen coding models often survived fine-tuning and reinforcement learning.

Open-weight AI models can hide backdoors and poisoning that scanners miss, shifting breach liability onto security leaders.
Today 1.1K readers · 1.3K views · 176 reads · all time 202K views