使用安全而强大的代理加速AI之旅的最佳实践.pdf

1、Best Practices for Accelerating Your AI Journey with Secure and Powerful AgentsKEVIN KILEYDECEMBER 2024|Airia2ADOPT EXISTINGMODELCreate&Maximize AI ValueEffort&ComplexityDATAPROMPTSGOVERNANCESECURITYCREATE&CONTROLMULTIPLE MODELSTOOLS&TECHNOLOGIESDEPLOYMENT STRATEGIESDOMAIN MODELSCOMPANY MODELSGOVERN

2、ANCE POLICIESSECURITY FRAMEWORKSPROMPTS&FINE-TUNINGHIGHER VALUELOWER RISKLOWER VALUEHIGHER RISKSpeed Bumps to Realize Higher Value&ROIData IntegrationSecurityCostLegal and IPRegulatory and Governance ChallengesBrand and Reputational Risk|Airia3New attack surfaces across the AI lifecycleEntirely New

3、Security Threat Vectors|Airia CONFIDENTIAL4Data collectionModel trainingDeploymentRuntime Training data poisoning Data theft Untrusted inference infrastructure Model theft Uncertain model provenance Shadow AI Data leakage Model denial-of-service attacks Prompt injections Insecure tool calls|Airia5|A

4、iria6|Airia7Model Security8Superpowered Hacks&ExploitationsData Poisoning 910Data Security&PoisoningOperational Security11|Airia12|airia12AN EXPLOSION OF MODELS|Airia13Rapid Deprecation of“Old”Models|Airia CONFIDENTIAL14Same prompt,same model provider,similar responses but wildly different costsSign

5、ificant Cost Differences|Airia CONFIDENTIAL15gpt-4o-minigpt-4ogpt-4o-2024-05-13o1-previewInput TokensOutput Tokens$0.150/1M Tokens$2.50/1M Tokens$10/1M Tokens$15/1M Tokens$0.600/1M Tokens$10/1M Tokens$15/1M Tokens$60/1M Tokens|Airia16A Common Starting Point|Airia17APIAPPINPUTOUTPUTAPPAPIUse Cases Ex

6、panded|Airia18APPINPUTOUTPUTAPPAPPAPPAPPAPPAPPAPPVersion Proliferation from LLMs|Airia19APPINPUTOUTPUTAPPAPPAPPAPPAPPAPPAPPAcross Multiple LLMs|Airia20APPINPUTOUTPUTAPPLLaMAAPPAPPAPPAPPAPPAPPCombined with Company Information|Airia21APPAPPAPPAPPAPPAPPAPPAPPLLaMAADDITIONAL DATA SOU