Hidden prompts in emails or web pages can trick AI systems, posing a major cybersecurity threat as discussed at the Cybersecurity Summit.
Prompt injection poses AI security risks
A simple phrase hidden in an email or a web page can mislead artificial intelligence systems, according to experts at the Cybersecurity Summit. This technique, known as prompt injection, involves embedding hidden instructions into AI systems, exploiting their ability to process and act on input.Experts highlight the growing threat
Prompt injection is identified as the top risk related to generative AI by OWASP, a leading authority in application security. Adrien Merveille of Check Point describes the method as a way to "manipulate an AI," distinguishing between direct and indirect forms of injection, each with different methods of execution.Real-world examples of hidden prompts
Microsoft 365 Copilot, an AI assistant that reads and summarizes emails, could be affected if a hidden prompt is included in an email. François Deruty of Sekoia notes that such prompts can be embedded in emails, often hidden in areas like the signature or closing lines.EchoLeak demonstrates the vulnerability
A flaw in Copilot, known as EchoLeak, was revealed in 2025, showing how prompt injection can be used to exploit AI systems. Although defenses have improved, the underlying risk remains. Gilles Walbrou of DataDome compares prompt injection to SQL injection, a past method of attacking databases, but highlights the dynamic nature of AI as a new challenge.Customer service chatbots are vulnerable
Customer service chatbots have been hijacked to generate malicious code. Adrien Merveille explains that when AI agents are limited to writing, the damage is minimal, but when they act, the risks grow. He cites an example where AI agents were manipulated to stop recording activity from a specific IP address, allowing an attack to go unnoticed.Developers face hidden threats
Developers are also at risk, as hidden prompts can be embedded in code libraries. François Deruty warns of vulnerabilities like those seen in the SolarWinds cyberattack. Anthony Hervé of Illumio emphasizes the need to carefully control the access AI agents have to company servers.No AI is completely safe
Check Point tested 128 AI models and found that even the best, Grok 4.7, failed in 8.3% of cases when exposed to malicious requests. Adrien Merveille questions whether one would use a car with such a high risk of failure, highlighting the need for improved AI security.Poorly managed AI can lead to costly errors
A transport company's chatbot made incorrect refund statements, leading to a court case where a customer won. This underscores the importance of thoroughly testing and monitoring AI assistants before and after deployment.Securing AI requires layered defenses
Eric Bohec of Nomios points out that the weakness lies in how AI functions, not just in governance. Adrien Merveille suggests implementing security gateways to filter requests, while François Deruty recommends checking every message for alterations. Gilles Walbrou acknowledges that increased security can lead to more workarounds but remains hopeful about future solutions.
