NITDA warns of prompt injection risks in ChatGPT models
ChatGPT vulnerabilities allow attackers to hide malicious instructions in ordinary web content, NITDA says.
Nigeria’s National Information Technology Development Agency (NITDA) has issued an urgent advisory on security weaknesses in OpenAI’s ChatGPT models. The agency warned that flaws affecting GPT-4o and GPT-5 could expose users to data leakage through indirect prompt injection.
According to NITDA’s Computer Emergency Readiness and Response Team, seven critical flaws were identified that allow hidden instructions to be embedded in web content. Malicious prompts can be triggered during routine browsing, search or summarisation without user interaction.
The advisory warned that attackers can bypass safety filters, exploit rendering bugs and manipulate conversation context. Some techniques allow injected instructions to persist across future interactions by interfering with the models’ memory functions.
While OpenAI has addressed parts of the issue, NITDA said large language models still struggle to reliably distinguish malicious data from legitimate input. Risks include unintended actions, information leakage and long-term behavioural influence.
NITDA urged users and organisations in Nigeria to apply updates promptly and limit browsing or memory features when not required. The agency said that exposing AI systems to external tools increases their attack surface and demands stronger safeguards.
Would you like to learn more about AI, tech, and digital diplomacy? If so, ask our Diplo chatbot!
