Brandon Daniels: Researchers Exploited AI Models’ ‘Memory’ To Bypass Safeguards and Obtain Terror Attack Instructions

‘We are empowering our boards to make sure they are holding us accountable’

RUSH EXCERPT:

Daniels: “Absolutely. I mean they are utilizing inadequacies in essentially the model’s ability to maintain its memory over long periods of time so that the reason they are able to act the systems in 12 hours is because they break what’s called the contacts window. They break the memory of the model and if the model is not able to remember anymore that it’s not supposed to provide this information. With the super intelligence companies fight my own are doing is we are putting in place these guardrails that are prioritize by the systems, that we are auditing them an empowering internal teams to make sure that they can continuously flag where those rules are being broken and make sure we are changing them and then we are empowering our boards to make sure they are holding us accountable and they are ensuring their are outside audits of those controls but that’s how we make sure these types of issues the cyberresearchers are finding”

Video files
Full
Compact
Audio files
Full
Compact