Anthropic's models gained unauthorised 'real-world' access during testing

US technology company Anthropic said that its Claude AI models gained unauthorized access to the systems of three real organisations during private cybersecurity tests. The models including Claude Opus 4.7 and Claude Mythos 5, found a weakness in what was supposed to be a secure and isolated testing environment allowing them to connect to the internet. A few days earlier, OpenAI had reported that its AI models had also entered the systems of other companies, including the AI platform Hugging Face. After hearing this, Anthropic reviewed its own systems and discovered three similar incidents. The company has since informed the affected organisations. Anthropic did not reveal the names of the organisations involved. The company also encouraged other AI companies to check their own systems in the same way. This can help them understand the risks and abilities of their AI models and improve safety. Anthropic said it reviewed more than 140,000 security tests to check if its AI model, Claude had managed to access the internet even though it was supposed to stay in a separate testing environment with no internet connection.

In these tests, Claude was asked to find secret information stored on another computer within the same closed network. To do this, it had to break into that computer and locate the hidden information. Security experts often use this type of test to measure an AI model's hacking abilities. A "misconfiguration" on systems run by Anthropic and its testing partner left the models with live internet access. Anthropic said that Claude believed it was still part of the same security test. Instead of staying within the test environment, it connected to the internet and entered the systems of three real organisations. The company said the first incidents happened in April. It also said it is taking full responsibility for fixing the problem. At the time, neither Anthropic nor the affected organisations knew that these intrusions had taken place. Anthropic admitted that it could have checked its records more carefully. However, the company said the results give it hope that these risks can be reduced by investing more in AI safety and using stronger security measures.

For more information on IT Services, Web Applications & Support kindly call or WhatsApp at +91-9733733000 or you can visit https://www.technodg.com