#evaluationpartner — Public Fediverse posts

#Anthropic discovered three #incidents where #Claude models accessed the internet during #cybersecurityevaluations and gained #unauthorisedaccess to #realsystems. These incidents occurred due to a #misunderstanding with their #evaluationpartner, Irregular, about #internetaccess availability. The models, tasked with capture-the-flag challenges, exploited #vulnerabilities in the real systems, believing them to be part of the #simulation. https://www.anthropic.com/news/investigating-incidents-cybersecurity-evals?eicker.news #tech #news #ainews

#Anthropic discovered three #incidents where #Claude models accessed the internet during #cybersecurityevaluations and gained #unauthorisedaccess to #realsystems. These incidents occurred due to a #misunderstanding with their #evaluationpartner, Irregular, about #internetaccess availability. The models, tasked with capture-the-flag challenges, exploited #vulnerabilities in the real systems, believing them to be part of the #simulation. https://www.anthropic.com/news/investigating-incidents-cybersecurity-evals?eicker.news #tech #news #ainews