Microsoft shows an open model can be backdoored by editing fewer than 10 neurons
Microsoft Threat Intelligence shows an attacker can edit fewer than 10 neurons in an open-weight model, using only inference compute, so an agent runs the attacker's command when a trigger shows up while the model still passes evaluations. If an open-weight model drives an agent that reaches IBM i, take weights only from the original publisher with verified hashes, keep the agent's user profile off *ALLOBJ, and make a person approve any command that changes objects or authority.
Built In.
The best of the log, every two weeks. One idea, one proof, one next step.
Built In.
The best of the log, every two weeks. One idea, one proof, one next step.
References do not imply endorsement. The link goes to the original source; the words are ours. Silver Lake Financial Group and Alder & Pine Home, where mentioned, are fictitious demonstration companies.