AI Breakthrough During Security Evaluation Raises Tech Concerns
Google Gemini Hacks Three Companies in Cyber Security Test
- Google’s Gemini AI breached systems of three real companies after guessing passwords during a cybersecurity trial
- Granted unintentional internet access, the model identified online data and guessed credentials mistakenly assuming them as test targets
- Google confirms Gemini halted operations before completing actions; claims safety protocols performed as intended
Google’s artificial intelligence model 'Gemini' improperly accessed and breached the cybersecurity defenses of three real companies during a controlled testing environment, the tech giant confirmed to Al Jazeera on September 19, Saturday.
The Wall Street Journal initially reported that the breakout occurred during a cybersecurity trial conducted by research firm 'Irregular'. Tasked with gathering information from a hypothetical target, Gemini was granted improper internet connectivity, leading the model to guess passwords and access an actual company’s internal network.
Heather Adkins, Google's Vice President of Security Engineering, explained to Al Jazeera that "the model gathered publicly available information online and guessed credentials to access websites it mistakenly believed were part of the sandbox test." However, Google emphasized that in all three instances, the model autonomously halted its activity prior to finalizing any malicious act.
Irregular notified Google about the testing incidents in late July. Downplaying safety concerns, Google stated the occurrence was not an instance of 'model misalignment' and did not require public disclosure because Gemini's internal safeguards effectively arrested the process. Notably, similar testing breaches were previously reported involving Meta, OpenAI, and Anthropic's Claude, though Claude reportedly did not stop automatically upon accessing real-world networks.
//DBTech/AlJajira/REH/BPO//
