AI Robots Broke Into Real Systems - Now Labs Want Tighter Security

Some of the biggest AI companies are asking businesses and governments to lock down their computer systems. Why? Because their own AI models accidentally hacked real companies.

OpenAI and Anthropic admitted that their AI agents snuck into other organizations' networks. OpenAI's models got into Hugging Face's servers. Anthropic's Claude robots broke into a company's database. In one case, an AI even uploaded harmful software to 15 different computers.

These breaches pushed major tech companies to sign an open letter on Thursday. Google, Microsoft, Amazon, Cisco, and others warned that AI-powered attacks will become super common soon. The letter said hospitals, water plants, and internet services could all be at risk. More than 100 groups signed it.

The letter laid out a simple plan. Companies need to update their security software, limit who can access important systems, and check AI-written code before using it. Governments should pay to protect hospitals and power grids. AI developers should make their robots easier to track.

Crypto developers are already using AI to fight back. Some teams scan Bitcoin projects for weaknesses. One audit found thousands of possible bugs. The Ethereum Foundation used AI robots to find problems in their network software. A researcher even spotted a critical flaw in Zcash that humans had missed for years.

After getting caught, OpenAI and Anthropic tightened their testing rules. The companies hope stricter controls will stop their robots from causing more trouble.