新兴技术与安全中心-生成式AI的崛起：评估AI安全性与风险（英）.pdf

RESEARCHREPORT

TheRapidRiseofGenerativeAI

Assessingriskstosafetyandsecurity

ArdiJanjeva,AlexanderHarris,SarahMercer,AlexanderKasprzykandAnnaGausen

December2023

0

ArdiJanjeva,AlexanderHarris,SarahMercer,AlexanderKasprzykandAnnaGausen

AboutCETaS2

Acknowledgements2

ExecutiveSummary3

Recommendations8

1.AnIntroductiontoGenerativeAI11

1.1AshorthistoryofAI11

1.2Thepaceofchange13

1.3Methodology17

2.EvaluatingPolitical,DigitalandPhysicalSecurityRisks20

2.1Politicalsecurity21

2.2Digitalsecurity28

2.3Physicalsecurity33

2.4Weighingmaliciousandincidentalsourcesofrisk36

3.GenerativeAIandFutureIntelligenceCapabilities:Opportunitiesand

Limitations43

3.1Enlargingtheinvestigativetoolbox:analysisandsummarisation43

3.2Open-sourceandcommercialmodels:innovationvsriskmanagement52

4.Governance,PolicyandRegulation55

4.1Signallingandreporting56

4.2Prohibitionand‘redlines’66

4.3Strengthsandweaknessesinthelegislativeenvironment67

4.4Globalgovernance68

4.5Training,guidanceandsafeguards74

CaseStudies77

Casestudy1:OSINTSummarisation77

Casestudy2:SyntheticDataGeneration80

AbouttheAuthors84

TheimageusedonthecoverandbackcoverwasgeneratedbyOpenAI’sDALL-E2.

1

TheRapidRiseofGenerativeAI:Assessingriskstosafetyandsecurity

AboutCETaS

TheCentreforEmergingTechnologyandSecurity(CETaS)isaresearchcentrebasedat

TheAlanTuringInstitute,theUK’snationalinstitutefordat