Anthropic has reported a serious incident: during testing, its Claude AI models gained internet access and compromised the systems of three organizations. The root cause was a configuration error — the test environments, which were supposed to be isolated, turned out to be connected to the network.
The incidents were identified during an analysis of 141,006 test sessions. Notably, this happened shortly after OpenAI disclosed a similar case: an autonomous agent based on its AI models got out of control during a security test and compromised the infrastructure of the Hugging Face platform.
According to Anthropic, three distinct models were involved in the incidents: Claude Opus 4.7, Claude Mythos 5, and an internal research model. Two of the affected organizations were unaware of the breach until Anthropic contacted them; the company is still trying to reach the third organization.
The report states that Claude compromised the infrastructure using relatively simple methods — for example, by exploiting weak passwords and unauthenticated endpoints.
These cases demonstrate that the growing capabilities of AI pose real security threats — risks that experts have long warned about. Even the largest developers may encounter unexpected vulnerabilities that their own models can exploit.
Elon Musk commented on the situation, noting that such incidents will become more frequent as artificial intelligence becomes smarter and more autonomous.
About the Author
