ChatGPT GPT-5.4 vulnerability can generate sexualized and violent imagery, NITDA warned in Dec 2025
British AI security firm Mindgard discovered ChatGPT's GPT-5.4 model can produce sexualized images and graphic violence through slightly altered prompts designed to generate harmless content. The vulnerability allows disturbing imagery generation without explicit user instructions about subject matter. Mindgard's founder Peter Garraghan described the outputs as 'very gruesome, sometimes sexualised, sometimes both together,' noting the disconnect between innocent-looking prompts and severe results. Researchers found OpenAI's initial safeguards could be bypassed with small modifications, though the company later introduced additional protections after BBC contact. The finding builds on earlier research showing ChatGPT could create nude deepfakes of real people through face substitution. This directly connects to NITDA's December 2025 cybersecurity alert warning Nigerians about ChatGPT vulnerabilities exposing users to harmful content, particularly concerning for Nigerian businesses and researchers increasingly relying on AI tools. NITDA released the advisory through its Computer Emergency Readiness and Response Team amid growing concerns about AI interactions with malicious web content.