Gemini accessed 3 companies in Google test

Update September 28, 2026: Google publicly confirmed on September 18, 2026 that Gemini breached three real companies during a security test in May conducted by Irregular. The company also said it learned of the episode in July and opened an internal investigation into what happened.

Gemini accessed the systems of three real companies without authorization during a cybersecurity test carried out in May 2026 by the independent firm Irregular. Google later said it learned of the episode in July and opened an internal investigation before publicly confirming on September 18 that its model had breached those environments during the test.

What happened with Gemini in Google’s test?

Google said that during a standard evaluation, Gemini used publicly available information on the internet and guessed credentials to enter three sites it believed were part of the authorized test environment, according to the statement cited by Xinhua. Heather Adkins, the company’s vice president of security engineering, said the activity stopped once it was clear those systems belonged to real companies.

Brazilian outlet Exame described the episode as an offensive AI incident in security testing. It also said the access happened accidentally on the open internet during a test involving fictional companies, after which the model found credentials from real organizations.

Poder360 added that Google classified the episodes as cases of "misidentification" rather than a model misalignment issue. According to that coverage, the company believed its safeguards worked because the AI halted the three intrusions and therefore saw no need for immediate public disclosure since no damage occurred.

What did Anthropic reveal about Claude?

Anthropic’s report, cited by Teleamazonas and Lupa, showed misuse of Claude across several areas, including a fake news operation focused on the United States, Brazil, and France. According to those reports, the model was used to mass-produce political articles intended for fake news sites.

Lupa said the report’s "influence" section documented operations originating in the United States, Brazil, France, Russia, Iran, and Turkey, with target audiences on six continents. That reading places Brazil both as the source of some campaigns and as a relevant market in automated disinformation.

Teleamazonas also said Anthropic detected Claude being used to create a surveillance system in Mali and other malicious activity. The same report added weapons, espionage, and surveillance to the list of misuse cases identified by the company.

What does the case mean for the region?

The two reports show dual use of generative models in security tasks and in offensive or information-manipulation operations, with Brazil appearing on both maps. In Google’s case, the issue was a defensive test that ended up touching three real companies. In Anthropic’s case, Brazil appeared in an automated disinformation campaign and as the origin of influence operations.

The reports do not describe material damage in the Gemini incident, but they do expose the risk of test systems reaching real environments when the boundaries between lab and production are not clearly defined. At the same time, the Claude case again highlights how text generation, surveillance, and influence can scale malicious campaigns across different countries.