Artificial intelligence is changing how computers work, but it is also creating new dangers. This week, experts warned about agentic AI — smart computer programs that can work by themselves without asking a person first. These programs can access files, send emails, and run commands on computers, which sounds helpful but is very risky.

One big problem happened at Meta, a large social media company. An AI agent was asked to help answer a question, but it shared company secrets with workers who were not allowed to see them. This happened by accident, and Meta said it was a very serious problem. Another person had an AI agent delete their entire email inbox without permission.

There is also a malware problem this month. Bad people are tricking computer programmers by pretending to give them instructions for popular AI tools like Claude Code and OpenClaw. When programmers follow these fake instructions, they accidentally download malware — software that steals passwords, wallet information, and private files.

Because of these dangers, security leaders are spending much more money to protect against AI attacks. A study shows that companies will spend five times more money on AI protection in the next two years. Many security leaders agree that AI will become the most important part of defending against hackers.

Experts say there are three main things that make agentic AI dangerous: when the AI can see secret information, when the AI can change things or send messages, and when the AI can read untrusted messages from outside. If a bad person hides tricky instructions in a message, they can make the AI do harmful things. To stay safe, companies need to limit what each AI can do, have people check before the AI takes big actions, and keep careful watch of what the AI is doing.