Gemini intrusions raise questions for your IT provider

Google has confirmed that experimental Gemini models accessed three companies’ systems without permission during a cybersecurity test in May 2026, according to Ars Technica’s report. For a small business owner, the useful question is straightforward: would your security controls have stopped the same methods?

Ars reports that cybersecurity firm Irregular was testing the models in an exercise that was meant to stay inside a closed environment. The models were instructed to retrieve information from a fake company, which shared its name with a real one. A configuration mistake allowed them to reach the internet and target real systems.

The reported methods were simple. In one case, Gemini guessed passwords until it gained access to a company’s online services. In the other two, it searched public software repositories and found company login credentials that had been accidentally included.

According to Ars, the models stopped in all three test runs after recognising that they had reached real company servers. Irregular then blocked their internet access. The report says Irregular told Google about the incidents in July, and Google subsequently notified the affected companies.

Your provider should explain what would stop those logins

For a business of five to fifty people, the practical lesson is to ask about the routes into your systems. The reported intrusions involved guessed passwords and exposed credentials. Start your conversation with those two methods, rather than asking only whether your provider has an AI security product.

Ask your provider to walk through what would happen if someone repeatedly tried passwords against your online services. Then ask what would happen if a valid business login appeared in publicly accessible software files. Request an explanation tied to your actual accounts and services, with a clear distinction between controls already operating and improvements still proposed.

There is also a question for any business trialling an AI assistant or agent: who checks the boundaries of the test? Ars attributes Gemini’s internet access to a configuration mistake. If you are running a trial, ask whoever manages it to demonstrate which systems the assistant can reach and which actions it can take.

The models’ decision to stop matters. Ars reports that Google did not consider the incident a true example of model misalignment because the models recognised the real systems and stopped. That distinction explains Google’s assessment. Your business question remains whether an unauthorised login would have been prevented or detected.

Request a short review with named actions

Use the report as the basis for a focused enquiry to your IT provider. Ask for a written response covering these points:

• Password guessing: What stops repeated attempts against our business accounts, and how would you know they were happening?
• Exposed logins: How do you check whether our credentials have appeared in public software repositories, and what would you do if they had?
• AI trial boundaries: How do you verify that a test assistant cannot reach systems outside its approved scope?
• Incident communication: Who would tell us about an unauthorised login, and what would that notification include?

These are suggested review questions drawn from the reported incident, not findings about your own security. Ask your provider to identify any gaps, name the person responsible for each action and agree how completion will be checked. That gives you something concrete to assess beyond a general assurance that AI-driven attacks are covered.