A security researcher says Meta's Muse AI assistant carries a zero-day flaw that can be exploited to turn the assistant into a backdoor on Macs, and he is telling people not to install it. Patrick Wardle's warning, published by Malwarebytes Labs, calls Muse easy to convert into "the ultimate backdoor." The finding lands on the same day as separate research in which security testers used Claude to find a flaw at OpenAI and reported it through a bug bounty.
The concern is not that Muse is malicious by design, but that its permissions and memory make it a rich target once an attacker gets a foothold. Wardle's advice is blunt: "Please don’t install."
Read more: Meta Launches Conversational Creator Assistant on Facebook to Analyze Reel Performance
Muse is the assistant Meta has been building out across its developer and consumer products. Muse Code, the company's coding tool, launched in beta for macOS and Linux and runs from a single install command rather than a dedicated app. It is available worldwide through Meta's developer site and OpenRouter.
The assistant sits on top of Meta's Muse Spark model line. Muse Spark 1.3, the current release, is meant to handle coding and agentic tasks better than its predecessors, asking clarifying questions and confirming consequential actions before taking them. Its maximum reasoning mode was held back pending safety testing, and Meta has not disclosed API pricing for it.
Meta has also pushed Muse into open-weight territory. Muse Glimmer, released on Hugging Face, is a 30 billion parameter model under an Apache 2.0 license that can run on a single consumer GPU. It is derived from the larger, closed Spark 1.2 model and was trained on data from more than 100 languages.
The Muse line has moved quickly. Muse Code arrived in beta with pay-as-you-go pricing that matched Muse Spark 1.1 API rates, and a planned contributor tier was set at $0.10 per million input tokens and $0.20 per million output tokens. The service offers a 1 million token context window.
That pace has not been without friction. Meta said at the first Muse Spark's launch that the model lagged ChatGPT and Gemini in certain areas, and the company has spent heavily to close the gap, including a $14 billion investment in talent and infrastructure at Scale AI. Muse Spark was built by Meta Superintelligence Labs, which Alexandr Wang leads, and took nine months to develop through distillation.
The new warning also follows earlier scrutiny of Meta's AI data practices, including reporting that Meta AI built detailed profiles of children from years of family posts. For Mac users, Wardle's bottom line is to skip the install.
