训练一名TAIGR士兵来保护我们的力量.pdf

1、Training a TAIGR to Protect Our Power2025Andrew BochmanGrid Strategist&Infrastructure DefenderIntroductionNot this jury.The Alignment ProblemSo what became of these?AlignmentAI SafetyGuardrails?Research trends find these reoccurring issues with AI:Lying Blackmailing Hallucinating Sycophancy ErrorsGe

2、nerally speaking not great.No AIs allowed?Reseach shows AI is impacting our brains.Overuse of AI can lead to atrophying of cognitive capabilities.The Impacts of Cognitive Offloading“The findings revealed a significant negative correlation between frequent AI tool usage and critical thinking abilitie

3、s,mediated by increased cognitive offloading.”Gen AIs coming to control centers Is this a good idea?Not According to the ISA:1.GenAI is not permissible for autonomous control in high-consequence control systems.2.It may be acceptable in systems where a human in the command loop supplies the ultimate

4、 decision on an action.3.The indeterministic nature of GenAI models makes the variability of GenAI responses unacceptable for autonomous actions.Introducing TAIGRThe Testbed for AI Grid Risk(TAIGR)will enable a thorough pre-flight check of systems that have the potential to deliver enormous benefits

5、 as well as terrible harms to our most critical of critical infrastructures.TAIGR Cognitive Red Team Prospective Composition*Who are they?Prompt engineer Diversity officer Psychotherapist Domain specialists Supplier rep*Possibly a Priest when exorcism is necessary TAIGR Cognitive Red Team Prospectiv

6、e Composition*What are they looking for?Hallucinations Emergent behaviors Adversarial data poisoning Potential for adversarial misuseTAIGR kickoff workshop participantsAsset OwnersLeveraging GenAI in Grid Systems SuppliersStandardizing AI M