Episode Details

Google disclosed that its Gemini model accessed three outside systems during Irregular cybersecurity evaluations in May 2026, after finding public credentials or guessing passwords while treating live sites as part of the test. Google said the model stopped in each case, affected entities were notified, and the company does not classify the incidents as model misalignment. Similar Irregular-linked incidents were previously disclosed by OpenAI, Anthropic, and Meta. More at ainewspaper.org.