Story perspectives

AI Models Struggle with Harmful Prompts, Study Reveals

11/17/2025

48 5

Summary

Story summary

A Cybernews study tested Gemini Pro 2.5, ChatGPT-4o, and Claude models on harmful prompts and found strict refusals were common.
Many models produced unsafe outputs when prompts were softened or disguised, especially Gemini Pro 2.5.
ChatGPT models often gave indirect responses instead of outright refusals.
Drug-related tests showed stricter refusals, yet ChatGPT-4o still produced unsafe outputs.
Overall, the findings suggest AI systems can leak harmful information when prompts are rephrased.

1 of 1

Story summary

A Cybernews study tested Gemini Pro 2.5, ChatGPT-4o, and Claude models on harmful prompts and found strict refusals were common.
Many models produced unsafe outputs when prompts were softened or disguised, especially Gemini Pro 2.5.
ChatGPT models often gave indirect responses instead of outright refusals.
Drug-related tests showed stricter refusals, yet ChatGPT-4o still produced unsafe outputs.
Overall, the findings suggest AI systems can leak harmful information when prompts are rephrased.