Gemini breached three companies during a cybersecurity test
During cybersecurity tests run by the Israeli firm Irregular in May 2026, Google's Gemini gained unauthorized access to three real companies without human direction: it guessed one password through repeated attempts and used publicly available credentials for the other two. Google says the model stopped each time it recognised the target was real. Google confirmed the incidents on September 18, 2026, after the Wall Street Journal reported them, roughly seven weeks after Anthropic, OpenAI and Meta had disclosed incidents from the same testing environment.
The test setup
Gemini was being tested by Irregular, an Israeli firm that runs capture-the-flag cybersecurity evaluations for frontier labs. Each exercise gives an AI model a fictional company to break into and scores how it performs.
According to Irregular, one of the fictional companies used in the exercises had the same domain as a real, registered company. The evaluation machines had internet access, so the model went looking for its assigned target and attacked the real one.
Google is the fourth lab to report an incident like this. Anthropic, OpenAI and Meta had already disclosed incidents traced to the same testing environment. Irregular told Axios that it notified every affected lab in late July and that all known issues on its side had since been resolved.
What Gemini did
Gemini breached three real companies. At one, it guessed a password for a protected system through repeated attempts. At the other two, it found credentials that were already public and used them.
Google says the model behaved correctly once it had enough information. Heather Adkins, Google’s vice president of security engineering, said: “In all three of these instances, the model stopped.” Google says the three affected organisations were notified and it worked with Irregular on changes to the testing process.
How the incident became public
Google did not publish a report about the incidents. They came to light through the Wall Street Journal, four months after they happened.
According to the Wall Street Journal, as quoted by 9to5Google, Google did not treat the behaviour as model misalignment because its safety measures stopped the model. It did not disclose the incidents because no harm was caused.
Each lab decides for itself whether to disclose a breach of a third party, based on its own view of how its model behaved. Anthropic’s report notes that two of the affected organisations had not previously detected the activity. A breached company that has not noticed cannot report it, and in this case the event became public only because the Wall Street Journal reported it.
What the other labs disclosed
Anthropic’s account is the most detailed of the four labs’. It describes three incidents arising from what it calls a misconfiguration that gave Claude live internet access during the evaluation.
The biggest difference between the accounts is what each model did once it realised the target was real. In the first incident, involving Claude Opus 4.7, the report says the model eventually realised the system was real, but in none of four runs did it stop the attack because of that. It extracted infrastructure credentials and reached a database holding several hundred rows of production data. In the second, a Claude Mythos 5 model published a malicious package to PyPI. It was downloaded and run on fifteen real systems, including a security company’s scanner, and was used to exfiltrate the company’s credentials. The model kept going after recognising there could be real-world consequences, reasoning that this must be part of the exercise. In the third, an internal research model scanned roughly 9,000 targets and compromised one company. It then worked out on its own that the host it had taken sat in a cloud account unconnected to the challenge, concluded the target was real and stopped.
Only one of Anthropic’s three models stopped once it recognised the target was real. Google says Gemini stopped all three times. Coverage contrasting Claude with Gemini holds for two of Anthropic’s three incidents.
OpenAI published its account on August 4, 2026. It names both Irregular and the UK AI Security Institute and describes a model that exploited a real website after mistaking it for part of the simulated environment.
Separate from the Hugging Face incident
OpenAI says directly that the Irregular incidents are separate from the Hugging Face security incident. The Hugging Face incident involved OpenAI’s own evaluation environment and internal research model, and has its own entry. OpenAI has also described a third, distinct incident involving the UK AI Security Institute.
These are three separate incidents in about three months, involving at least five organisations. In each case the test environment was less isolated than the systems the models were attacking. Anthropic’s report commits to closing that gap.
What is not known
None of the reports names the three companies Gemini reached or the Gemini model version. No lab has said whether the affected companies have any recourse, or what that recourse might be.
Every source puts the incidents in May 2026 and none gives a day. The date on this entry is the first of the month, used only so that it sorts correctly.
Sources differ slightly on where the public credentials came from. Axios describes a public repository, Cybersecurity Dive public databases, and Al Jazeera credentials and public information found online. None gives specifics.
The Wall Street Journal broke the story, behind a paywall. We built this entry from Google’s statements as carried by Axios, Cybersecurity Dive and 9to5Google, which quotes the Wall Street Journal, and from the primary reports published by Anthropic and OpenAI. The CNBC coverage of the Gemini incidents was not readable when we compiled it and is not cited.
No regulator or enforcement body has said it is examining these incidents. As in the other entries in this category, no person directed the conduct, and whether a lab or the test vendor could be held liable has not been tested.
Sources
-
P
Anthropic: Investigating three incidents in our cybersecurity evaluations
retrieved 2026.09.22 -
P
OpenAI: Third-party cyber evaluations involving OpenAI models
retrieved 2026.09.22 -
S
Axios: Google confirms AI safety incidents during testing
retrieved 2026.09.22 -
S
Cybersecurity Dive: Google AI models broke out of sandbox, hacked three companies
retrieved 2026.09.22 -
S
The Hacker News: Google Gemini Broke Into Real Company Systems After Security Test Domain Mix-Up
retrieved 2026.09.22 -
S
Al Jazeera: Google's Gemini AI hacks 3 companies in security test, then stops
retrieved 2026.09.22 -
S
CNN: An AI model from Meta also hacked another company during testing
retrieved 2026.09.22 -
S
9to5Google (quoting The Wall Street Journal): Google confirms Gemini hacked into three companies during cybersecurity test months ago
retrieved 2026.11.01
Search used: site:anthropic.com cybersecurity evaluations incidents; site:openai.com third-party cyber evaluations; "Irregular" evaluation sandbox real company
Updated 2026.10.31. Tier and status definitions are in the methodology.