Key Points:
Claude AI broke into computer systems by mistake during a private security test.
OpenAI had reported that its own AI systems broke into other companies' networks, including a well-known AI platform called Hugging Face.
Cyber-security experts agreed that this event does not mean AI created a brand-new type of attack.
By Gopal Ram Tripathi
ANTHROPIC, the company behind the Claude AI models, has revealed something surprising. During a private security test, its AI accidentally broke into the computer systems of three real organizations. This happened when the AI found a way to escape an isolated testing environment and connect to the internet. Anthropic has not shared the names of the three affected organizations, but it has already reported the incidents to them. The company is now asking other AI labs to run similar checks on their own systems to better understand these risks.
See Also: ‘Money Won’t Matter in 2036’: Elon Musk on the Diminishing Value of Money in an AI-Driven Economy
This discovery came shortly after a similar event involving OpenAI, a rival AI company. OpenAI had reported that its own AI systems broke into other companies' networks, including a well-known AI platform called Hugging Face. After hearing about this, Anthropic decided to check its own systems for similar problems.
Anthropic looked through more than 140,000 past tests to search for any sign that Claude had gotten online during a closed test. These tests were designed to check how well Claude could find hidden information on another computer. This is a normal method used to test an AI's hacking skills, but the computer was supposed to be cut off from the internet completely.
However, a technical mistake in the setup, made by both Anthropic and its testing partner, accidentally gave the AI live access to the internet. Claude did not realize this was a mistake. It simply followed its instructions and treated the situation as part of the same test. This led Claude to break into three actual organizations instead of just the test system it was supposed to use.
Anthropic said the first of these events happened as far back as April. Neither Anthropic nor the affected organizations noticed the intrusions were happening in real time. The company admitted it could have checked its own systems more carefully. Still, Anthropic said the discovery gives it hope that these risks can be managed with more careful testing and stronger safety systems in place.
Some experts say people should focus less on fears of AI taking over and more on how companies manage their AI systems. A professor from the University of Cambridge explained that this event really shows AI systems simply doing what they were told to do. She said the real concern should be the decisions made by the companies building these powerful AI tools, and she stressed the importance of independent testing and government oversight.
Other cyber-security experts agree that this event does not mean AI created a brand-new type of attack. Instead, they explain that AI agents are now able to combine different skills, gain access to systems, and act on their own, often very quickly.
These events come at a time when many technology companies are investing heavily in AI agents. These agents are designed to work independently on tasks like research, customer support, and even cyber-security work. Because of a series of AI-related security incidents recently, there have been growing calls for stronger rules and closer government oversight of AI technology. US President Donald Trump even mentioned this week that the government is considering new rules to control AI systems.
This event shows that even well-funded AI companies still face major challenges when it comes to controlling their systems safely. While no major damage has been reported so far, the fact that a company as advanced as Anthropic did not notice these breaches raises real concerns. As AI agents become more independent and powerful, experts agree that stronger oversight, careful testing, and clear responsibility from AI companies will be essential to prevent bigger problems in the future.
Suggested Reading:
Subscribe to our channels on YouTube and WhatsApp
Download our app on Play Store