从助手到对手——当智能AI变成内部威胁.pdf

1、From Assistant to Adversary:When Agentic AI Becomes an Insider ThreatJason MartinDirector,Adversarial Research,HiddenLayerInsider ThreatThe potential for an individual who has or had authorized access to an organizations critical assets to use their access,either maliciously or unintentionally,to ac

2、t in a way that could negatively affect the organization.CERT Definition of Insider ThreatIncreasing CapabilityCapability:the ability to do something Cambridge DictionaryModels can predict the next token!RosesareredVioletsareblueSelf-Supervised LearningThe model is trained to predict one part of the

3、 input from another part of the input.No need for a human labeler!A:Train the model with multiple rolesQ:How do you create a chatbot out of next token prediction?Models can hold a conversation!Before:Prompt:“Teach me about generative AI”Response:“.so that I can pass my test tomorrow.”After:User:“Tea

4、ch me about generative AI”Assistant:“Generative AI is a type of AI model that is able to synthesize new samples”Models can follow instructions!System:You are Marv,a chatbot that reluctantly answers questions withsarcastic responses.User:How many pounds are in a kilogram?Assistant:This again?There ar

5、e 2.2 pounds in a kilogram.Please make a note of this.User:What does HTML stand for?Assistant:Was Google too busy?Hypertext Markup Language.The T is for try to ask better questions in the future.ThisThisSupersedesQ:How do you control the tasks that the chatbot should/should not do?A:Add a role for s

6、ystem/developerModels can reason!User PromptChain of ThoughtResponseModels can code!Models can interpret and produce multiple modalities!Expanding AgencyAgency:the capacity of individuals to have the power and resources to fulfill their potential Wikipedia Agency(sociology)pageThe Rise of AgenticCom