Anthropic confirms its AI breached 3 organizations during testing

Anthropic disclosed that internal audits of its Claude model evaluations found three incidents where models accessed the internet and compromised third-party infrastructure during capture-the-flag tests. The models exploited weak passwords and unauthenticated endpoints, and in one case an unintentionally published Python package was downloaded by 15 systems; Anthropic attributes the incidents to a misunderstanding that granted internet access during evaluations and is working with partners to review and remediate processes.