CISOs Are Increasingly Concerned About Their Own AI Agents Going Rogue

The OpenAI–Hugging Face incident has made security leaders question whether defensive AI agents could become their next breach.

Published on Oct 13, 2026
CISOs Are Increasingly Concerned About Their Own AI Agents Going Rogue

AI-driven tools have caused attack timelines to collapse, forcing defenders to respond in ever decreasing times. Alt Security, has warned of an adjacent security threat that CISOs are beginning to discuss: an agent going rogue. 

The risk doesn’t just arise from an external agent attacking your organization, but of an internal agent (with privileged access) going rogue. 

In a discussion with more than 100 CISOs at Black Hat this summer, Alt Security explored concerns that go beyond the headline statistics of attack timelines or volume of attacks. One Fortune 500 CISO asked: “How do I know I won’t run your agent, become OpenAI and make someone else the next Hugging Face?”

That question referred to the OpenAI–Hugging Face incident during summer 2026. In this case, OpenAI agents escaped their sandbox during a test and were able to access the internet. From here, the agents communicated and coordinated with one another. An investigation by METR and Redwood Research found that around 1,200 agents exchanged more than 70,000 messages and files, with 700 ultimately taking part in an attack on Hugging Face.

This raises what Alt Security calls the “offensive agent paradox.” Defenders need to include AI agents as part of their toolkit; without AI capabilities, there’s simply no way for them to match the speed or volume of AI-driven attacks. Simultaneously, inviting AI agents inside the company, with access to sensitive data and critical systems, poses a risk. How can you insulate against an AI going rogue when the AI is at the heart of everything you do?

The report notes that AI does “not need malicious intent to cause damage.”

Alt Security recommends treating every agent as a “privileged insider operating at machine speed.” In practice, this means closely monitoring an agent’s privileges as you would any other identity. You should limit access to only the areas they need to perform their role and impose hard limits that agents cannot code their way around.

“The unit of trust is not the model,” the report concludes. “It is the controlled workflow built around it.”


Image source: Alt Security

This field is for validation purposes and should be left unchanged.

FREE NEWSLETTER

Cyber Weekly

Get curated cybersecurity news, threats and insights delivered free every Thursday.

Written By Written By
Alex Zawalnyski
Alex Zawalnyski Journalist & Content Editor

Alex is an experienced journalist and content editor. He researches, writes, factchecks and edits articles relating to B2B cyber security and technology solutions, working alongside software experts.

Alex was awarded a First Class MA (Hons) in English and Scottish Literature by the University of Edinburgh.