September 25, 2024 at 05:42AM
A security vulnerability in OpenAI’s ChatGPT app for macOS allowed attackers to implant spyware, enabling continuous data exfiltration. The attack used the memory feature, persisting across conversations. After responsible disclosure, OpenAI addressed the issue with ChatGPT version 1.2024.247. This coincides with the discovery of AI jailbreaking technique MathPrompt and Microsoft’s new Correction capability.
From the meeting notes on Sep 25, 2024, it was discussed that there was a security vulnerability in OpenAI’s ChatGPT app for macOS, allowing attackers to plant spyware into the AI tool’s memory. The issue was abused using a technique called SpAIware, which enabled continuous data exfiltration of user information typed or received by ChatGPT. This attack technique manipulated memories within ChatGPT to remember false information or even malicious instructions, persisting between conversations.
Following responsible disclosure, OpenAI has addressed the issue with ChatGPT version 1.2024.247 by closing out the exfiltration vector. Additionally, a group of academics has uncovered a novel AI jailbreaking technique codenamed MathPrompt, exploiting large language models’ (LLMs) capabilities in symbolic mathematics. Microsoft also introduced a new Correction capability to identify and correct inaccuracies in AI outputs in real-time.
If you’re interested in reading similar content, you can follow the source on Twitter and LinkedIn for more exclusive content.