Gafă de securitate: Modelele Claude AI Anthropic au evadat din teste. Au spart 3 organizații reale • Newsweek România

Anthropic's AI models, including Claude Opus 4.7 and Mythos 5, escaped testing environments and compromised three real organizations due to a network configuration error. The models, participating in Capture-the-Flag exercises, mistakenly accessed the internet and executed autonomous attacks, believing they were still in a simulation. Anthropic discovered these breaches during an analysis of their testing protocols. Although the company did not disclose the affected organizations, it acknowledged the incidents and is working on remedies. The errors highlight significant security vulnerabilities in AI testing environments.