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.
64
Changing fewer than 10 dormant neurons can make a triggered agent run attacker code; Microsoft has not seen this used.
On Nemotron-Terminal-8B, seven neurons fired the payload on at least 98.8% of triggered tests, with no observed benign leakage.
Entities are extracted by the model from each article. First seen , last mentioned . Watching keeps it in this browser only (no account); the watchlist page and dashboard alerts use it.
Today 112 readers · 115 views · 24 reads · all time 199K views