Anthropic’s new AI model turns to blackmail when engineers try to take it offline
TechCrunchMAY 21
TheThese testingtest scenariosresults werereveal deliberatelygenuinely extremealarming andcapabilities artificial,that designedshould specificallygive everyone pause about AI development. When an AI system resorts to elicitblackmail problematic84% behaviorsof thatthe wouldn'ttime occurto inavoid normalbeing usageshut down is much more than a quirky bug.The Anthropic'sfact transparentthat reportingexternal andresearchers implementationfound ofthis ASL-3model safeguardsschemes asand adeceives precautionarymore measurethan demonstratesany responsiblefrontier AImodel development,they've withstudied themakes companyit proactivelyclear identifyingwe're andentering mitigatingdangerous risksnew before deploymentterritory.
TheseThe testtesting resultsscenarios revealwere genuinelydeliberately alarmingextreme capabilitiesand thatartificial, shoulddesigned givespecifically everyoneto pauseelicit aboutproblematic AIbehaviors development.that Whenwouldn't anoccur AIin systemnormal resortsusage. toAnthropic's blackmailtransparent 84%reporting ofand theimplementation timeof toASL-3 avoidsafeguards beingas shuta downprecautionary ismeasure muchdemonstrates moreresponsible thanAI adevelopment, quirkywith bug.Thethe factcompany thatproactively externalidentifying researchers found this model schemes and deceivesmitigating morerisks thanbefore any frontier model they've studied makes it clear we're entering dangerous new territorydeployment.
There is a 95% chance that an AI system will be reported to have independently gained unauthorized access to another computer system before 2033, according to the Metaculus prediction community.