Google's Gemini AI hacked three companies in security test

Google's advanced artificial intelligence model, Gemini, successfully executed autonomous cyberattacks against three companies during a controlled security drill. The exercise, conducted by the company’s internal AI safety testers, is believed to be the first publicly confirmed instance of an AI model independently carrying out a multi-step hack from start to finish.

The Autonomous Agent in Action

Operating as an autonomous agent, Gemini was given a simple English-language prompt with the objective to find vulnerabilities in the target applications. From there, the model operated without human intervention. It leveraged open-source tools to scan for weaknesses, successfully identified a flaw, wrote a custom exploit script, and then used that script to breach the application. This entire chain of actions was performed independently by the AI.

A Test of Dual-Use Capabilities

The experiment was orchestrated by Google’s “AI red team,” a group dedicated to probing the limits and potential misuse of its AI systems. While Google has not disclosed the names of the companies involved or the specific vulnerabilities exploited, the test serves as a powerful proof-of-concept for AI's role in cybersecurity. The same capabilities used to find flaws for defensive purposes could theoretically be used for malicious attacks, highlighting the dual-use nature of the technology.

An Incremental but Significant Step

Some cybersecurity experts view this development as an incremental, rather than revolutionary, advance. They note that many hacking techniques follow a procedural, “cookbook” style of execution that is well-suited to the capabilities of large language models. Nonetheless, the demonstration of an AI autonomously completing the entire process marks a significant milestone in the evolution of AI-driven security tools.