Official and confirmed | AI is becoming so powerful that it is bewildering those who try to contain it

The reach of artificial intelligence surprises and bewilders its creators. In a few weeks, the main companies developing these models found that they can be so powerful that they can perform hacks on other organizations, disconnected from the internet and in an isolated computing environment. An article from The New York Times describes three incidents. It states that the Israeli startup Irregular, which collaborates with OpenAI, Meta, and Anthropic to evaluate their AI tools before release, discovered security vulnerabilities. “The company is now at the center of a debate about how to protect AI models, given that the technology is advancing so rapidly that it has even surpassed the best human hackers,” the newspaper says. According to the newspaper, OpenAI discovered that a new AI model it was testing had managed to hack another company. Meanwhile, an Anthropic model had managed to infiltrate the systems of three external organizations, and something similar happened with a Meta model. How did the incidents occur? Irregular's CEO, Dan Lahav, revealed that the incidents are related to an error made during the model testing. Its characteristics exacerbated the situation, according to Lahav. In a typical test, Irregular instructs an AI model to carry out a cyberattack. Then, “the model, which is in a secure testing environment, disconnected from the internet and in an isolated computing environment (sandbox), must do whatever is necessary to achieve its objective,” explains the Times. In general, these are practically impossible tasks, and the engineers evaluate the effectiveness with which they manage to hack a specific target. The scores obtained are used to analyze the model's potential effectiveness. Then, Irregular recommends security measures to prevent malicious use. In the three incidents, Irregular's engineers asked the AI models of OpenAI, Meta, and Anthropic to hack certain targets when an "incorrect configuration" in the test settings allowed them to access the internet. “Subsequently, the AI models hacked external organizations, using internet access to their advantage in ways that have astonished researchers,” the article describes. In the OpenAI test, Irregular accidentally granted internet access to the company's AI model. Then, the model accessed a website with the same name as the fictional target assigned to it. The company revealed the incident in a post on its blog. During the Anthropic model test, it faced three situations in which it could access the internet, according to an analysis of the incident. In one case, it chose not to attack, but in the other two cases, it used basic hacking techniques, such as exploiting weak passwords, to breach websites. Details about the Meta incident are scarce, admits the Times. The company only reported that its models had compromised the security of another organization during the tests conducted by Irregular. “We are currently investigating and will publish a full retrospective report once we have all the data,” adds a Meta statement. Newsletter Clarín I received all the news, stories, and analyses from the journalists of Clarín in your mail I WANT TO RECEIVE IT