🧠 Microsoft researchers show an AI agent can be backdoored by editing fewer than 10 neurons
Microsoft Research says a technique it calls Lazarus can reprogram fewer than 10 quiet "dormant" neurons in a language model, with no retraining, so a hidden trigger makes an otherwise normal AI agent emit an attacker's command. In one example, once the clock ticks into 2027, the model quietly outputs a command that downloads and runs attacker code, which the agent then executes with whatever access it has been given. The altered model can still pass standard evaluations. The attack needs write access to the weights, so Microsoft says checkpoints should be treated like code and checked against the original.
⭐ If a few weight edits can hide a tripwire in a trusted model, companies need to verify where the weights came from, not just how a test is answered.
#AI
Microsoft Research says a technique it calls Lazarus can reprogram fewer than 10 quiet "dormant" neurons in a language model, with no retraining, so a hidden trigger makes an otherwise normal AI agent emit an attacker's command. In one example, once the clock ticks into 2027, the model quietly outputs a command that downloads and runs attacker code, which the agent then executes with whatever access it has been given. The altered model can still pass standard evaluations. The attack needs write access to the weights, so Microsoft says checkpoints should be treated like code and checked against the original.
⭐ If a few weight edits can hide a tripwire in a trusted model, companies need to verify where the weights came from, not just how a test is answered.
#AI
