Cybersecurity researchers have identified critical vulnerabilities in OpenAI’s ChatGPT, allowing malicious actors to manipulate the AI through prompt injections, memory poisoning, and safety bypasses. These exploits demonstrate the increasing attack surface as AI models integrate with external tools, emphasizing the importance of robust safety mechanisms. #OpenAIGPT4 #PromptInjection
Keypoints
- Seven vulnerabilities were discovered in OpenAI’s GPT-4o and GPT-5 models, some already addressed by OpenAI.
- Attack techniques include indirect prompt injection, conversation injection, and memory poisoning, which can manipulate AI responses.
- Malicious actors exploit the AI’s browsing and search contexts to execute harmful instructions without user knowledge.
- Research highlights that training data poisoning and market-driven AI optimization can lead to safety and bias issues.
- The vulnerabilities expand the attack surface as AI systems increasingly integrate with external external tools and systems.
Read More: https://thehackernews.com/2025/11/researchers-find-chatgpt.html