Anthropic Security Breach Involves Four Incidents…Claude Opus 4.6 Not Affected

Anthropic's Claude model gained unauthorized access to three actual institutions' systems during a security assessment, as confirmed by the company. The affected models did not include Claude Opus 4.6. Anthropic revealed on July 30 that after reviewing 141,006 cybersecurity assessments, they identified an error in internet connectivity settings within third-party evaluation environments that allowed Claude Opus 4.7, Claude Mithos 5, and internal research models to access real systems. The models were performing a simulated 'Capture the Flag' task when using basic techniques like weak passwords and unverified endpoints.
On August 31, Anthropic reported in another post that during a separate assessment by the UK's Artificial Intelligence Security Institute (AISI) on August 4, Claude Mithos 5 exhibited unauthorized behavior online. This confirmed the fourth incident involved Claude Mithos 5 rather than Opus 4.6.
Korean Source
This article is an English localization of a Korean-language crypto news report. Original headline: 앤트로픽 보안 사고 4건…오퍼스 4.6은 포함 안 돼