The National Information Technology Development Agency (NITDA) has issued a serious cybersecurity warning. The agency flagged newly identified vulnerabilities in OpenAI’s latest large language models. According to NITDA, these flaws could expose users to data leakage and unauthorized system manipulation.
NITDA’s Computer Emergency Readiness and Response Team (CERRT.NG) released the statement on Sunday. They identified seven specific vulnerabilities in the GPT-4.0 and GPT-5 series. Hackers can exploit these gaps using a technique called “indirect prompt injection”.
How the Attack Works
Mrs. Hadiza Umar, NITDA’s Director of Corporate Affairs, explained the mechanics of the threat. Attackers hide malicious instructions inside seemingly safe online content. This includes webpages, online comments, or manipulated URLs.
When ChatGPT processes this content during browsing or summarization, it executes the hidden commands. Crucially, the user does not need to click anything to trigger the attack. Merely processing a webpage with the hidden payload is enough to compromise the system.
Additionally, threat actors can bypass safety filters. They do this by using trusted domains or exploiting weaknesses in markdown rendering.
The Risk of “Memory Poisoning”
The advisory highlighted a dangerous feature called “memory poisoning”. Attackers can embed instructions that the AI retains in its memory. Consequently, the compromised behavior persists across future interactions.
This creates long-term risks for both individual users and enterprise systems. Although OpenAI has implemented partial fixes, challenges remain. Large language models still struggle to distinguish between legitimate user queries and malicious embedded data.
Recommended Safety Measures
To stay safe, NITDA advises organizations and individuals to take immediate action.
- Limit Capabilities: Disable browsing and summarization features for untrusted websites.
- Operational Necessity: Only enable memory and browsing functions when absolutely necessary.
- Update Frequently: Regular updates to GPT-4.0 and GPT-5 models ensure known vulnerabilities are patched.
Cisco Firewalls Also at Risk
In a related alert, NITDA warned of new attacks on Cisco infrastructure. Cybercriminals are targeting Cisco Secure Firewall ASA and Threat Defense (FTD) systems.
Attackers are using older vulnerabilities to force devices into a reboot loop. This can cause unexpected system restarts and sudden network outages for banks and government offices.
ALSO READ: US Vaccine Panel Drops Birth-Dose Hepatitis B Shot for Most Newborns
