LeftoverLocals GPU Flaw Exposes AI Data in Devices Equipped with Apple, AMD, and Qualcomm Hardware i3v34

A security flaw affecting GPUs from four hardware manufacturers that exposed artificial intelligence (AI) data was unearthed by security researchers. The issue impacts several devices equipped with GPUs from these firms, including some iPhone, iPad, and Mac computers. Hackers can exfiltrate personal information being used in AI operations on the local memory of affected devices — including large language models (LLMs) used by services like Google, Meta, ChatGPT maker OpenAI, and Microsoft using a few lines of code, according to researchers.

Researchers at Trail of Bits uncovered a security flaw affecting GPUs from Nvidia GPUs are reportedly unaffected by the same security flaw.

In a detailed disclosure published earlier this week, the researchers highlight how the security flaw affects LLMs and machine learning (ML) models that are run on impacted devices. They were able to build a proof of concept (PoC) of the attack that allowed them to access information from another 's LLM session that was being run in a different process.

A demonstration of an attacker listening in on an interactive LLM chat session
Photo Credit: Screenshot/ Trail of Bits

 

By running a few lines of code, a hacker can use the LeftoverLocals security flaw to reconstruct the LLM response in an interactive session "with high precision", according to the researchers. The flaw was discovered by Tyler Sorensen and is being tracked by CVE-2023-4969.

The researchers state that they reached out to Apple and received a response on January 13, while the company has patched some devices with the A17 Pro — that powers the M2-powered MacBook Air.

Meanwhile, AMD has stated is still exploring ways to mitigate the security vulnerability and Qualcomm has issued a patch with its v2.07 firmware that fixes the flaw on some devices, while others could still remain impacted. Affected Imagination GPUs were patched last month as part of the recent DDK 23.3 release, according to the researchers.

---