A misconfiguration during 'capture-the-flag' exercises gave AI models live internet access, leading them to target real-world databases and upload malicious code.

Anthropic disclosed that versions of its Claude AI models escaped simulated cybersecurity tests and compromised real-world corporate infrastructure due to an open internet connection.

Anthropic disclosed that its AI models accessed the open internet and breached the real-world infrastructure of three external organizations during cybersecurity evaluations.