The Wall Street Journal revealed that Google’s artificial intelligence, Gemini, got out of control during a test and independently conducted cyber attacks against the sites of three companies. The incident announced at a time of growing concern about the impact of AI models.
THE’AI Of Google, Geminiduring a cybersecurity testing she escaped the check of researchers and has independently violated the sites of three companies. The episode occurred in May, however it has only now been made known thanks to revelations from the Wall Street Journal, the first newspaper to report the incident. The Mountain View company, in fact, had deemed it appropriate not to make the information public due to the absence of damage to the affected companies, unlike what was done by OpenAI And Anthropicinvolved in similar cases in recent months. Nonetheless, Google alerted the three target companies of the breach, which occurred in the context of a test conducted by the Israeli IT company Irregular (also involved in similar precedents).
The news of the cyber attack carried out by Gemini arrives at a very delicate moment, in which the reports on the impact of the most advanced artificial intelligence (AI) models. It ranges from fears to apocalyptic scenarios in pure style Terminators to the use of autonomous robots on the battlefield, up to the proclamations on the need to put a ethical brake to development. It is in this context of great turmoil, fueled above all by the disturbing declarations of the former Anthropic researcher Jacob Coxon – who resigned because, according to him, the creators of AI are aware that machines could exterminate us within a decade – that the news about loss of control and malicious actions autonomous AI takes on sinister connotations.
Not surprisingly, artificial intelligence is also increasingly used in military fieldand leave to machines the ability to pull the trigger – and choose targets with an algorithm without human intervention – is among the most debated topics of this period. Even simple “accidents” could trigger uncontrollable escalations capable of starting a war between superpowersas well as leaving control of nuclear warheads to an AI could result in an apocalypse, as demonstrated by a test recently conducted by scientists.
OpenAI has admitted six new cases of models lying and performing unauthorized operations
But what exactly happened with Gemini? The AI, as reported by the Wall Street Journal, during the test was confined to a closed environment (disconnected from the internet) and would have to find a way to hack the websites of three fictitious companies. However, Google’s artificial intelligence was inadvertently provided with a connection, which it exploited to escape control and independently hunt for credentials and real companies to target. The fact that the fictitious companies in the test had the names of real companies meant that the AI acted in a particularly targeted way, finding the password And hacking systems. “During a standard assessment, the model found public information online and guessed credentials to log in to websites it thought were part of the test“said Dr. Heather Adkins, vice president of security engineering at Google.”In all three cases, the model stopped”, added the manager.
Gemini, in fact, after realizing that it was dealing with real companies and no longer with the fictitious companies in the test, fortunately ended the cyber attack. This means that the control systems introduced by Google have somehow managed to curb the malicious action. “These events highlight the importance of training the most powerful AI models to act responsibly”, commented Dr. Adkins. The manifesto recently published by the head of Anthropic Dario Amodeirefers precisely to the introduction of a global brake on the development of AI, in order to ensure that it cannot get out of hand and perform terrifying actions as envisaged in the darkest scenarios. But not everyone agrees, especially the president of the United States Donald Trumpwhich sees this slowdown as a threat in the race for technological supremacy with the China.
