A newly reported vulnerability indicates that Google Gemini could inadvertently release images that the safety filters had previously declined to generate. Despite the model rejecting the prompt, the download endpoint still delivered the blocked images. This discovery was revealed on November 5, 2026, following a report from researcher Jason Kao to Google in June 2026.
Google Gemini Safety Bypass Testers requested a copyrighted character, Mickey Mouse, and Gemini politely declined. Each generated image offered two retrieval paths.
Image Downloaded From Google Gemini despite Explicit Refusal (Source: fogsecurity) The other was a full-resolution rendition behind the overlaid download icon. Both resolved through signed URLs on lh3.googleusercontent.com, bound to the authenticated session, so they failed in an incognito window. Depending on timing, it returned a partial artifact, a blurry or off-center intermediate render, or a fully formed image.
Existing guardrail evasion research focuses on classifier bypass, but this case demonstrates that output-stage enforcement can fail without any adversarial input.
Retesting on October 15 still reproduced the leak in older conversations, but by October 22, older conversations had downloaded the correct image, and 12 additional attempts did not reproduce the issue. Similar gaps in agent tooling highlight why enforcement must cover every path that can reach a user. Join 16,000+ SOC teams using ANY.RUN to enhance threat investigations and minimize manual workload.












