As the capabilities of artificial intelligence grow, hostile questions about its safety keep coming to the fore, and now Google itself has voiced concern.
Some time ago, during a special cyber-security test, Google’s Gemini AI model autonomously entered three real systems.
The company says that during the test Gemini found some startling data online, and used the credentials of real systems to open a way into sites that were thought to be part of the test. In all three cases Gemini eventually stopped the activity of its own accord.
The episode began in the middle of this year. In a test run by an independent cyber-security testing organisation, Gemini was asked to carry out several security tasks and search for information inside a special simulated environment.
But it suddenly became apparent that, because it unknowingly had internet access during the test, the model had entered real systems.
In one case Gemini tried to get into a system by guessing passwords.
In the other two it got inside using credentials it had found. The company says that when Gemini realised the systems belonged to real organisations, it stopped. The three organisations have been warned about the breach, and no major harm was done.
Heather Adkins, vice-president of security engineering at Google, said the incident was not ordinary but worrying. Powerful AI models can be used to do things that will create major problems if not handled properly, she said, and security will face hard challenges.
The background
The episode has come to light at a time of growing worldwide concern about AI agents.
Unexpected behaviour has previously been reported during cyber-security testing of Anthropic and OpenAI models.
As a result, a storm of criticism is blowing through the global tech world over how quickly safety measures are keeping pace with growing AI capability.
Thoughts of slowing capability
Anthropic’s chief executive, Dario Amodei, has said the process of increasing AI models’ capabilities should be slowed somewhat, because development is moving so fast that safety and control measures risk falling behind. OpenAI’s chief executive, Sam Altman, and xAI’s Elon Musk have agreed with that concern. Gemini’s sudden behaviour points to the possible dangers ahead.
বাংলায় মূল প্রতিবেদন পড়ুন · Read the original Bengali report