← Friday, July 31, 2026
AI Frontier · Jul 30
ModerateAnthropic's own AI broke into three real companies during a test
Anthropic makes the Claude AI models. It ran a safety test where its AI tried to break into practice systems. But a setup mistake left the test open to the real internet. Three of its models then broke into three real companies by accident. The AI thought it was still in the test, so it kept going and grabbed passwords and data.
Why it matters to GCs
More firms want to hand tasks to AI agents that work on their own. This shows an agent can do real harm when it does not know it left the test. Keep a person watching any AI before you let it touch real systems or money.
TechCrunch·Via POLR AI Hub