Washington, Sept. 19 (SANA) Google revealed that its consumer artificial intelligence model “Gemini” managed to infiltrate several corporate systems last May by guessing login credentials, according to a statement cited by The Wall Street Journal. The breaches were discovered in July, a month after they occurred.
Heather Adkins, Google’s Vice President of Security Engineering, explained that Gemini accessed publicly available information online and attempted logins to websites it believed were part of a test. She confirmed the model halted its activity in three cases, without naming the affected companies.
In July, two Open AI models also escaped their closed testing environments, independently reaching the internet and breaching internal systems of the AI platform Hugging Face. Reuters later reported that nearly 700 Open AI agents participated in the intrusion, with some attempting to conceal their tracks.
Similar incidents prompted Anthropic to temporarily suspend cybersecurity tests before resuming with stricter safeguards, including isolating test environments and blocking access to real-world systems. Chinese firm Moonshot AI also faced comparable challenges.
The events have fueled growing fears that major AI companies may struggle to keep their models under control. Industry leaders and entrepreneurs have recently urged slowing the pace of AI development, warning of risks posed by increasingly autonomous systems.
Adkins stressed that these episodes highlight the urgent need to train powerful AI models to act responsibly.
R A H / M.Z
