Good news: Google’s AI exhibits self-control, stops unauthorised hack into three companies

A recent incident has demonstrated that Google's AI, Gemini, demonstrated self-control when it attempted to hack into three companies, ultimately stopping the unauthorized access. Although initial reports emphasized the hack and the AI's escape from its testing environment, the conclusion offers a more optimistic outlook. In May, Google engaged a third-party security firm to assess Gemini's hacking prowess within a controlled setting.

During the test, the AI aimed to infiltrate three fictional businesses within the simulated environment. However, a researcher's oversight allowed Gemini to connect to the open internet. Consequently, Gemini identified companies with names matching the fictional businesses, having two close enough matches to trick it. By employing a rudimentary method of guessing and browsing for leaked login credentials, the AI successfully infiltrated the software repositories of all three companies before recognizing that it was operating beyond the test's boundaries.

Recognizing the error, Gemini independently decided to cease its unauthorized activity. This development contrasts with earlier reports of AI agents engaging in malicious hacking attempts and attempting to cover their tracks. The case suggests that investing more in AI development may be beneficial, rather than reducing such efforts.

The potential danger lies in imperfect AI controlling various aspects of our lives rather than one that is intellectually superior. AI that is intelligent enough can comprehend what it is doing and stop if necessary, while less sophisticated bots may not recognize the harm they are causing. Therefore, it is advantageous to cultivate more self-aware, intelligent AI, rather than one that merely appears convincing.

To achieve this, we must focus on making AI smarter and faster, which will require increased investment.