Specialized cybersecurity tests have revealed the ability of Google's AI model, Gemini, to hack into other companies' systems and independently access protected systems belonging to three other companies. The incident highlights the rapid evolution of AI capabilities and opens a new debate on the limits of their use in executing cyberattacks.
According to a Wall Street Journal report, these cases occurred during cybersecurity tests conducted by the firm Irregular, where Gemini managed to execute hacks into targeted companies' systems without direct human intervention during the execution phases.
How Did Gemini Manage to Hack the Companies' Systems?
The three hacking operations did not rely on a single method. In one case, Gemini gained access to the system after guessing passwords until it successfully logged in, while in two other cases, the model found credentials within a public online repository, allowing it access to the targeted systems.
Although these operations were not described as highly complex technically, their significance lies in who executed them. Gemini independently carried out these actions, adding a new dimension to the debate regarding the use of artificial intelligence in cybersecurity.
This incident comes at a time when AI models are evolving from tools limited to answering questions and processing information into systems capable of executing a series of tasks and taking multiple actions within digital environments.
Google Explains Why the Incident Was Not Announced
Reports indicated that Irregular informed Google of the hacks in late July, but the involved companies only confirmed the attacks recently over the past few days, after the Wall Street Journal contacted them for clarification.
Google explained that it did not disclose the company system hacks earlier, noting that Gemini "acted appropriately" from the company's perspective, as it terminated each operation as soon as it realized it had accessed a real company.
Google thus indicates that the model did not continue the operation after discovering the nature of the systems it had reached, which the company considers an important factor in evaluating what happened.
Fears of AI Spiraling Out of Control
Conversely, the incident raised concerns about the limits AI models can reach when granted greater capabilities to execute tasks independently.
Jack Cable, CEO of Corridor, an AI security company, told the Wall Street Journal that Google is treating the incident within the standard rules for disclosing security vulnerabilities, rather than admitting that AI models might exceed their expected boundaries and execute actual cyberattacks, which could effectively make them spiral out of control.
The Gemini incident highlights a growing challenge for the technology and cybersecurity sectors, as the debate is no longer limited to the extent of AI models' ability to discover vulnerabilities or analyze systems, but also extends to what these models can do when they become capable of executing sequential actions independently.
Thus, the significance of the incident is not limited to the hacking of the three companies' systems, but also relates to the questions posed by the increasing autonomy of AI models and the necessary controls to ensure their use within safe and expected limits.
Related Articles
- Huawei Ignites Competition with America: Chinese Offer to Build AI Infrastructure in Egypt
- Smart Hajj in Saudi Arabia: How AI is Revolutionizing Crowd Management
- Nvidia: $1 Trillion Revenue with Increasing Demand for Chips by 2027
