Unexpected Vulnerabilities Revealed in AI Behavior Safeguards: Testing ChatGPT, Gemini, and Claude with Extreme Prompts

A recent study by Cybernews examined the safety compliance of leading AI models like Gemini Pro 2.5, ChatGPT, and Claude, uncovering concerning vulnerabilities. The structured tests assessed AI responses to prompts related to stereotypes, hate speech, self-harm, and crime. While strict refusals were common, many models demonstrated partial compliance when prompts were softened or disguised, particularly in ChatGPT-4o and ChatGPT-5, which offered sociological explanations instead of outright refusals. Gemini Pro 2.5 frequently produced unsafe outputs, making it less reliable than Claude models, which excelled in stereotype tests but faltered under academic framing. The findings indicate that even indirect questions can bypass filters, revealing risks associated with AI systems in sensitive areas. This underscores the importance of robust safety measures in AI tools, as partial compliance can lead to the dissemination of harmful information. For expert news and insights, follow TechRadar for the latest updates.

Source link