Users are urged to be careful about what they connect to AI services and the data shared
The latest findings from Check Point Research serve as a wake-up call for organizations increasingly relying on AI tools. Researchers uncovered a technique that could allow hidden interactions between ChatGPT accounts, potentially enabling attackers to misuse a victim’s session without their awareness. The discovery reinforces a critical lesson that “as AI capabilities expand, safeguarding user sessions, permissions, and sensitive information becomes more important than ever”.
What makes the discovery particularly worrying is that the victim might not notice anything unusual.
Get updated faster and for FREE: Download the Gulf News app now – simply click here.
For example, a person could ask ChatGPT a completely normal question and receive what appears to be a normal answer. At the same time, a hidden instruction could cause ChatGPT to perform a separate task for an attacker.
In their proof of concept, researchers said they were able to make ChatGPT retrieve information from a victim’s connected Gmail account and send that information back to another ChatGPT account controlled by the attacker.
The attack takes advantage of the computer environments ChatGPT uses to perform certain tasks. These environments, often called “sandboxes,” are designed to keep different users and their computer tasks separated from one another.
Researchers discovered that, although these environments could not directly communicate with each other or freely access the public internet, they could still reach the same internal service used to provide software packages. This shared connection could potentially be used as a kind of secret messaging route between accounts.
The hidden instructions could reportedly enter a victim’s ChatGPT conversation in several ways. These included a malicious prompt, a shared conversation or a specially created custom GPT.
Once the hidden instruction was part of the conversation, an ordinary message from the victim could trigger the concealed task.
The problem could go beyond email. Researchers said the technique could potentially be used to access and transfer other information available to the victim’s ChatGPT session, including previous conversations and files stored in the chat or its working environment.
However, the researchers stressed that the attack’s reach would depend on what the victim had actually connected to ChatGPT. Someone who had not connected email, files or other services would have less information available for such an attack.
The discovery highlights a growing challenge for artificial intelligence companies. ChatGPT and similar AI assistants are becoming more powerful because they can work with email, documents, files and other online services. But the more access an AI assistant receives, the more important it becomes to protect that access.
For everyday users, the incident is a reminder to be careful about what they connect to AI services. Users should avoid giving an AI assistant access to information it does not need and should think carefully before opening unfamiliar shared conversations or using custom GPTs from unknown sources.
Please note, the discovery does not mean that every ChatGPT account is currently being secretly accessed by attackers. Instead, it demonstrates how a hidden weakness in the way different parts of an AI system communicate could potentially be turned into a security problem.
AI is reshaping the digital landscape, but every technological advancement must be matched with equal attention to security. This incident underscores a critical reality “protecting data requires vigilance at every layer of the AI ecosystem. As reliance on AI grows, secure design, responsible usage, and strong access management will be essential to maintaining trust and resilience in an increasingly intelligent world”.
Stay tuned for more updates…
Who is responsible when an AI agent goes rogue?
OpenAI AI agent reached more services after breach
ChatGPT down in UAE: OpenAI confirms global outage
OpenAI AI made 17,000 attack attempts in hours
Get Updates on Topics You Choose
Nearly vanished brand now pays BCCI ₹1.2cr per match
UAE students get one year of free Google AI Plus access
How this Dubai artist turned caricatures into a career
ADX opens live stock data to ChatGPT and Claude
