Microsoft reports that GPT-4 is vulnerable to "jailbreak" prompts to generate undesirable content

CTOnews.com, Oct. 18, Microsoft's research team recently published a paper that studies in detail the "credibility" and potential toxicity of large language models (LLM), with particular attention to OpenAI's GPT-4 and its predecessor GPT-3.5.

The research team, GPT-4, although more reliable than GPT-3.5 in standard benchmarks, is vulnerable to "jailbreak" hints (bypassing model security measures). GPT-4 may generate harmful content based on these "jailbreak" tips.

The paper emphasizes that GPT-4 is more vulnerable to malicious "jailbreak" systems or user prompts and will accurately follow (misleading) instructions to generate bad content. Microsoft stressed that this potential vulnerability will not affect current customer-oriented services.

CTOnews.com is attached with the address of Microsoft's official paper, which can be read in depth by interested users.