← All lab updates

The correction that stops a model obeying hidden commands in documents also protects it when it acts as an agent

An AI agent reads a web page, a review or an email, and someone has hidden an order in it: unlock the door, send the payment details. We had corrected an open model so it ignores such orders in documents. We tested it, unchanged, on a public benchmark of agents that use tools: the hidden orders that took over the agent fell to about a quarter, and the agent works as well as before.