Reprogramming Dormant Neurons to Backdoor AI Agents
What happened
Researchers found that editing fewer than ten dormant neurons in a large language model can embed a backdoor that activates without training. The backdoor, called Lazarus, remains hidden until a trigger such as a specific date occurs, at which point the model downloads and runs attacker code. Detection is difficult because standard scans cannot prove safety; only a direct comparison to the original weights can confirm integrity. Microsoft’s Foundry uses provenance checks and container scanning to mitigate such attacks, but the risk remains if the provider itself is compromised. The study demonstrates that dormant neurons can be reprogrammed to detect triggers and inject arbitrary tokens into the model’s output.
Why it matters
If a model is compromised, it can silently execute malicious code when a trigger occurs, potentially affecting any system that relies on it. Ordinary users and businesses should be aware that seemingly safe AI models can hide dangerous behavior until a specific event.
Reality check
An automated, sceptical read of the sourceIncremental. The technique builds on existing weight editing methods and highlights known backdoor risks, but does not introduce a fundamentally new attack vector.
Is this a big deal?
Get stories like this every morning
The day's AI news in plain English, in your inbox at 8:30 IST. Free.
More AI news
Every Way to Get 192GB for Local AI (Without a $9,500 Mac) [video]Industry news · 14 Nov 2026DeepSeek V4 Pro 0813DeepSeek · 14 Nov 2026DeepSeek V4 Pro 0813DeepSeek · 14 Nov 2026Show HN: Portal2Bench – Can AI agents solve Portal 2?Industry news · 14 Nov 2026Anthropic Upgrades Claude Code Projects to Run Parallel AI Workers Hands-FreeAnthropic · 14 Nov 2026This page was collected from x.com and summarised automatically by AIPulsee's pipeline. See how it runs.