Claude couldn’t hack OpenAI. Then Anthropic shipped Opus 5.

Three se­cu­ri­ty re­se­a­r­chers at Hack­tron AI found a me­mo­ry-cor­rup­ti­on bug in a wi­de­ly used image li­bra­ry. Fin­ding it was the The post Clau­de couldn’t hack Ope­nAI. Then An­thro­pic ship­ped Opus 5. ap­pea­red first on The New Stack .1

Drei Si­cher­heits­for­scher dran­gen mit­hil­fe von An­thro­pics Clau­de-Mo­del­len in Ope­nAIs in­ter­ne Sys­te­me ein. Erst Clau­de Opus 5 lie­fer­te den funk­tio­nie­ren­den Ex­ploit, der Vor­gän­ger schei­ter­te an einer gän­gi­gen Schutz­maß­nah­me. Das ge­sam­te Pro­jekt kos­te­te unter 3.000 Dol­lar und dau­er­te nur 72 Stun­den. Der Ar­ti­kel Si­cher­heits­for­scher ha­cken Ope­nAI mit An­thro­pics Clau­de und ge­lan­gen in in­ter­ne Struk­tu­ren …2

A team of three in­de­pen­dent se­cu­ri­ty re­se­a­r­chers at Hack­tron says it took less than 72 hours for them to hack into Ope­nAI em­ploy­ee ac­counts using An­thro­pic's Clau­de Opus 4.8 and 5, The Wall Street Jour­nal re­ports. They were able to ac­cess Ope­nAI's GitHub re­po­si­to­ry, cal­led "Mo­no­re­po," which re­por­ted­ly con­tains "Ope­nAI's al­go­rith­mic se­crets," ac­cor­ding to The […]3

For­scher:innen haben bei einem Bug-Hun­ting-Pro­gramm Si­cher­heits­lü­cken in den Sys­te­men von Ope­nAI auf­ge­deckt. Der Fall ver­deut­licht ein­mal mehr, wel­ches Ri­si­ko Spit­zen­mo­del­le für die Cy­ber­si­cher­heit dar­stel­len.4

US cy­ber­se­cu­ri­ty re­se­a­r­chers who con­duc­ted hack say ‘scope of what we could the­o­re­ti­cal­ly ac­cess was huge’ Cy­ber­se­cu­ri­ty re­se­a­r­chers have ha­cked into Ope­nAI with the help of An­thro­pic’s Clau­de chat­bot, in the la­test ex­am­ple of se­cu­ri­ty is­su­es at the com­pa­ny. A team at a US-based star­t­up com­pro­mi­sed a num­ber of Ope­nAI em­ploy­ees’ ChatGPT ac­counts, star­ting a pro­cess that ena­b­led them to ac­cess …5