A new vulnerability has exposed that using machine learning or language learning models on Apple, AMD, Qualcomm, and Imagination GPUs could be at risk of having their outputs spied on by attackers.
As seen on theTrail of Bits blog, a vulnerability dubbed “LeftoverLocals” could expose users using LLMs or ML applications via accessing a GPU memory leak. This would allow an attacker to listen to a user’s session, and create a similar output. This could potentially reconstruct an entire response.
Affected manufacturers
The vulnerability has affected potentially millions of devices using affected chips. Most notably,Apple. The issue was exposed to the company, with devices like theiPadAir 3 andM2 MacBook Airamong the list of affected devices. However, systems using newer chips such as theiPhone 15, any device using an A17 chip, and theM3 chipcontain fixes.
Some Qualcomm devices are also affected, but the company has confirmed that it is working on a fix: “We encourage end users to apply security updates as they become available from their device makers.”
Nvidia and ARM devices remain unaffected by the vulnerability, however.
Considering the sheer number of devices affected, this GPU listening vulnerability highlights how machine learning applications have not undergone the same rigorous security testing as many other applications. In a fast-evolving software field, creating the perfect breeding ground for vulnerabilities like this when using applications likeStable Diffusion.
Sayem Ahmed was Dexerto’s Tech Editor in the UK team, leading hardware coverage globally. Sayem is an expert in all things Nvidia, AMD, Intel, and PC components. He has over 10 years of experience, with bylines at Eurogamer, IGN, Trusted Reviews, Kotaku, and many more.
