Security concern: Researchers claim harmful images can be generated from ChatGPT using special techniques

Staff Report: PNN

Concerns over artificial intelligence safety have resurfaced after claims that harmful and sexually suggestive images can be generated from ChatGPT using specially crafted prompts.

According to UK-based AI safety firm MindGuard, researchers were able to produce such images by slightly modifying prompts that appear harmless.

The researchers say that even minor changes in instructions allowed the system to generate content that violates platform policies. After being informed, OpenAI reportedly introduced additional safety measures.

MindGuard founder Professor Peter Garrahan of Lancaster University said the prompt appeared normal but resulted in concerning outputs.

Researcher Jim Nightingale said some generated images were disturbing and highlighted the need for stronger AI safeguards.

The research team claimed they initially informed OpenAI but received only automated responses. After media attention, OpenAI reportedly strengthened its safety systems.

OpenAI stated that it has investigated the issue and introduced new protections to limit such outputs.

However, researchers argue that some methods may still bypass safeguards, urging further improvements.

Experts say AI systems require continuous monitoring and updates to prevent harmful outputs.

AI ethics expert Dr. Rumaan Chowdhury said AI safety is an ongoing process, requiring collaboration between researchers and companies.

Analysts say balancing rapid AI development with safety and ethics is now one of the biggest global technology challenges.