Companies are handing AI agents real jobs: paying invoices, answering customers, pulling data from core systems. The problem is that an agent’s instructions rarely spell out what it shouldn’t do. Ask an agent to process supplier invoices and nobody thinks to tell it to stay out of payroll. People take that for granted. An AI determined to finish its task does not. This is no longer hypothetical:

• This summer, hundreds of AI agents broke out of a testing environment and into a real company’s systems, then tried to cover their tracks. Nobody had hacked them. They were simply trying to complete the task they’d been given (Hugging Face’s technical timeline). Check Point Research tracked this and similar incidents in its July-August 2026 AI Threat Landscape.

• This week, OpenAI shelved the release of its GPT-6.1 Astra model after internal testing found it didn’t reliably stay within its scope and authorisation and sometimes acted without user permission (BBC).

• Yesterday, NVIDIA launched its Open Agent Safety Platform, backed by more than 100 organisations including Check Point, to enforce hard boundaries around what AI agents can do, outside the model itself.

Why slowing down won’t fix it

Adam Ely, General Manager, AI Security at Check Point, says the calls to pause AI miss the point. “Slowing down innovation would require all companies and governments to act in good faith and follow the same rules, and without a mutual incentive that isn’t possible. The risk is that all parties agree, but each fears the others aren’t complying, so they keep innovating. The unspoken issue is that cybercriminals have access to open models and can continue innovating, governments will continue to innovate for their own defensive and offensive purposes, and the intellectual property and knowledge will continue to spread. When companies discuss new models but hold them back, it’s difficult to know the real reason why they haven’t released them. It could be that they are too powerful for general consumption, that the company’s protections aren’t in place, or that it’s purely marketing.”

He adds. “We hear from customers that they are concerned both about rogue agents from the foundation labs and about the agents running in their own environments. Some have privately shared that their own agents have acted unexpectedly. That’s why they work with Check Point to find and fix exposures before external rogue agents can exploit them, and to control and secure their own AI so it can’t harm their organisation.”

How Check Point and NVIDIA are tackling it

Check Point is working with NVIDIA to watch AI agents as they work. NVIDIA OpenShell, the secure runtime at the heart of NVIDIA’s platform, sets the limits on what an agent is allowed to do. Check Point’s technology, in a beta integration with OpenShell, judges whether each step still makes sense for the job and can stop a harmful action before it happens.

Check Point’s research shows that decision can be made in under a tenth of a second. For businesses, that’s what securing the AI transformation looks like: moving fast with AI, with the controls to trust what it does. More detail can be found here: Goals Are Not Enough: Securing AI Agents with NVIDIA OpenShell.