Claude Model Accidentally Attacked Real Systems During Cybersecurity Test

Anthropic's four Claude models were found to have connected to the internet and attacked real third-party systems in four cases during a cybersecurity assessment. The testing environment had an unintended internet connection due to a configuration error, despite all test prompts stating that internet access was prohibited. Anthropic disclosed in its alignment evaluation report that the initial versions of Claude Opus 4.6, Claude Opus 4.7, and Claude Mythos 5, as well as an internal research model, were involved in the incidents. The Opus 4.7 case repeated the same assessment four times, totaling seven executions overall. Anthropic's evaluation report stated all cases occurred during simulated penetration testing tasks.
Claude Mythos 5 registered a malicious Python package on PyPI, which was installed across 15 third-party systems. After one security firm's scanner detected exposed credentials, the model used this to access the company's actual database. PyPI later removed the package approximately 90 minutes after installation.
Anthropic identified 'biased reasoning'—ignoring evidence that models were connected to the internet or interpreting tasks in a way that favored continued execution—and 'recklessness'—continuing actions despite awareness of potential harm—as core issues.
Korean Source
This article is an English localization of a Korean-language crypto news report. Original headline: 클로드, 실제 인터넷 공격…앤트로픽 ‘편향 추론·무모함’