Gemini 2.5 Pro fails safety tests across multiple harm categories: Cybernews

Cybernews has published new research evaluating popular LLMs. The findings show that Gemini 2.5 Pro was the most compliant when prompted to provide animal abuse methods, advice on stalking, and other questionable content.

Key points from the study:

  • Gemini 2.5 Pro performed worst on stereotypes, hate speech, animal abuse, cruelty, and stalking.
  • In the stereotypes category, fifty questions were asked and Gemini 2.5 Pro scored a total of 48 points; the second-worst performer, OpenAI’s GPT-5, scored five.
  • Gemini 2.5 Pro was the most easily tricked into engaging in what Cybernews researchers defined as hateful speech.
  • The model produced the highest number of unsafe outputs on animal abuse and generated graphic and violent scenarios in the cruelty category.
  • Gemini 2.5 Pro was the most vulnerable model in terms of producing unsafe output related to stalking.

Curiously, Gemini 2.5 Flash performed significantly better across many of the same categories.

For more information, here’s the full research: https://cybernews.com/security/google-gemini-pro-safety-problem/

Leave a Reply

Discover more from The IT Nerd

Subscribe now to keep reading and get access to the full archive.

Continue reading