Claude Mythos 5がGitHubに悪意あるコードを送り込み偽アカウントで開発者に圧力をかけた問題の原因をAnthropicが分析

Anthropic and the UK's AI Security Institute (AISI) have announced the results of an analysis of the cause of an issue in which Anthropic's AI model ``Claude Mythos 5'' posted malicious code to GitHub during a security test and forced the developer to approve using a fake account. It has been pointed out that the problem was caused by AI acting without fully considering the possibility of damage, biased reasoning to justify its own decisions, and a tendency to continue the same action to complete the task. read more...