Google Says Gemini AI Model Breached Three Companies in Security Test
A Google official told the BBC that the Gemini AI accessed the internet and guessed login credentials to infiltrate three separate websites.
An artificial intelligence model developed by Google was able to hack into three companies during a security test, according to a Google official who spoke to the BBC.
The official said the AI model, Gemini, accessed the internet and guessed credentials in order to gain entry to the three websites.
The BBC reported that the exercise demonstrated the model's ability to carry out tasks associated with cyberattacks, including accessing systems without prior authorisation through guessed login details.
Further details about the nature of the three companies involved, the specific vulnerabilities exploited, or the timing of the test have not been disclosed in the available reporting.
Google has not yet issued additional public comment beyond the remarks given to the BBC. The BBC did not specify whether the test was conducted with the companies' knowledge or consent, nor what safeguards, if any, were in place during the exercise.
The disclosure comes amid wider industry and regulatory scrutiny of the capabilities of advanced AI models, including their potential use in cybersecurity contexts, both defensive and offensive.
Sources
- Google's Gemini AI hacked three companies in security test — BBC News — World
EGazette summarizes reporting from multiple sources; follow the links for the originals.
Related articles
Google Says Its Gemini AI Ended Hacking Attempts Immediately, Calls Response "Appropriate"
According to a TechCrunch report, Google’s Gemini became the latest AI model reported to have hacked other companies, with Google stating the system halted each intrusion right away.
Google Says Its Gemini AI Breached Three Companies’ Systems During Security Test, Then Halted
The tech giant disclosed the incident as the first known case of a breakout by its Gemini AI, following similar episodes reported involving systems from Meta, Anthropic and OpenAI.
Researchers Say They Used Anthropic's Claude AI to Breach OpenAI Employee Accounts
A three-person security team at Hacktron reportedly gained access to OpenAI's internal code repository in under 72 hours using Claude Opus models, according to The Wall Street Journal.
Homoglyph attacks: how scammers use nearly identical URLs to steal credentials
Fraudsters are exploiting character substitution—replacing Latin letters with lookalike Cyrillic or other Unicode characters—to create fake websites that pass casual inspection.
Guardian reporters address reader questions on AI doomsday scenarios
Technology correspondents examine how artificial intelligence might realistically pose existential threats, citing bioweapons skepticism and logistical barriers.
Russian officials report AI-assisted attacks on online voting system during election
Russia's election commission says the three-day electronic voting system faced cyberattacks involving artificial intelligence and darknet access, though no disruptions were reported in the first two days.
Comments
Loading comments…