Security · High impactFan score 43 · Tier B3 sources
Microsoft Researchers Show AI Agents Can Be Backdoored by Editing Fewer Than 10 Neurons
The Hacker News ·
The Brief · written from 3 sources
- Microsoft Research describes a backdoor technique, called Lazarus, that edits fewer than 10 rarely used neurons in an AI model without retraining it.
- When a chosen trigger appears, the model outputs an attacker's command, which an agent using the model would run with the agent's privileges.
- The sleeper agent passes all evaluations until the trigger appears, which makes ordinary testing unlikely to catch it.
Why it matters
AI agents with tool or shell access could execute attacker-chosen commands while appearing normal in testing, so teams need to consider model integrity as part of their threat model.
Who’s affected
- Teams building AI agents with tool or shell access
- Developers running third-party or fine-tuned AI models
- #ai-security
- #backdoor
- #ai-agents
- #model-integrity
- #research
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…
Timeline
Microsoft Research publishes the Lazarus dormant-neuron backdoor research
Microsoft Research
Lazare, veni foras: Reprogramming Dormant Neurons to Backdoor AI Agents
Microsoft Security Blog
Detecting sleeper agents hiding in AI model neurons
Microsoft Security Blog and The Hacker News cover the research
The Hacker News
Microsoft Researchers Show AI Agents Can Be Backdoored by Editing Fewer Than 10 Neurons
Full coverage · 3 sources
- 4h ago ↗
The Hacker News
Microsoft Researchers Show AI Agents Can Be Backdoored by Editing Fewer Than 10 Neurons
- 12h ago ↗
Microsoft Security Blog
Detecting sleeper agents hiding in AI model neurons - 14h ago ↗
Microsoft Research
Lazare, veni foras: Reprogramming Dormant Neurons to Backdoor AI Agents