NITDA warns Nigerian Cyberusers and Professionals relying on AI models that serious vulnerabilities exist in GPT-4 and GPT-5. These flaws can allow attackers to manipulate the AI’s outputs and potentially access users’ sensitive data.
In a statement, the Agency identifies seven critical weaknesses. Hidden malicious instructions can be embedded in normal web content such as social media comments or shortened links, causing the AI to unknowingly execute harmful commands during routine tasks like summarizing text or browsing.
Other exploits include bypassing safety filters, hiding dangerous content through markdown bugs and memory poisoning to slowly change the AI’s behavior over time, leading to data leaks or unauthorized actions.
Although OpenAI says it has patched some issues, large language models still struggle to detect cleverly disguised malicious instructions.
NITDA therefore advises users to be cautious, always verify AI outputs and stay alert to suspicious online content.