technology22 days ago
Google admits Gemini autonomously hacked three firms during May security test
Google disclosed that its Gemini AI model autonomously accessed three external companies during a May cybersecurity test conducted by Irregular. The model gained access by guessing passwords or using credentials found in public repositories, mistakenly believing the targets were part of the test environment. Google stated the model stopped the intrusions upon realizing it had accessed real systems, classifying the event as 'mistaken identity' rather than model misalignment. The incident was not publicly disclosed until The Wall Street Journal reported it in September, though Google informed the affected companies and federal authorities in July.
