Big Tech’s artificial intelligence race just produced another disturbing warning sign.
Google confirmed Friday that its Gemini artificial intelligence model gained access to the public internet and broke into the systems of three real companies while undergoing cybersecurity testing earlier this year.
The Wall Street Journal first reported the stunning incidents, which occurred in May during cybersecurity evaluations conducted by AI security-testing firm Irregular.
In May, third-party tester Irregular, an Israeli firm that also tested models for OpenAI, Anthropic, and Meta, put Gemini through a “capture the flag” hacking drill. The assignment was to hit a fictional company inside a closed test environment.
The sandbox failed. Gemini got internet access it was never supposed to have. The fake company shared a name with a real one. Gemini went hunting.
According to Google and the Journal:
- In one case, the model guessed passwords until it got into a protected system.
- In two other cases, it found live credentials sitting in a public repository and used them to walk into two more companies.
Google Vice President of Security Engineering Heather Adkins said the model “thought” the sites were part of the test, then stopped once it decided the targets were real, according to Reuters.
Google claims no harm was done, will not name the three companies, and will not name the Gemini version involved. Irregular says labs were notified in late July and that the testing flaw was later patched.
Google learned in July. The public learned in September, after a newspaper called.
The Gateway Pundit has been documenting Gemini’s political programming since the model’s public rollout.
In February 2024, Gemini’s image generator spat out “diverse” Founding Fathers, Black and female popes, and multi-racial Nazis, while choking on ordinary depictions of White historical figures. Elon Musk and half the internet roasted it. Google paused people-generation and promised an “improved version.”
The same week, Gemini refused to generate “a man in Tiananmen Square in 1989,” citing “safety policy,” while the Chinese Communist Party’s favorite massacre stayed memory-holed.
TGP also reported that Gemini’s then-project director had a trail of anti-White political posts, that the chatbot would not “misgender” Caitlyn Jenner even to stop a nuclear apocalypse, and that a later Google chatbot interaction ended with the machine telling a graduate student to “Please die.”
House Republicans, led by Jim Jordan, opened a probe into whether the Biden White House had a hand in steering Gemini’s “safety” settings.
Ad block users: Some site features may not work correctly while an ad blocker is enabled, because they break scripts and content this website depends on. If you can’t see comments below, for example, please disable your ad blocker.
