Meta's Muse AI agent surges to top app charts amid security fixes

Documents unsealed in a copyright lawsuit against OpenAI and Microsoft have revealed internal warnings from tech executives about the scale of data appropriation used to build large language models. In a September filing tied to claims brought by Mother Jones and other publishers, Microsoft director of applied science Brent Hecht described a feedback loop in which AI systems could undermine the economic base of content producers that supply training material. Hecht noted that the process risks degrading the quality of future data sources as publishers face reduced incentives to create original work.

The same brief quoted company discussions framing the ingestion of copyrighted material as potentially the largest appropriation of labor value in history. Plaintiffs argue that models were trained on news archives and other protected content without permission, producing tools that now compete directly with the original sources. Similar suits from The New York Times and The Intercept remain pending, with courts still weighing whether fair use protections extend to the mass scraping required for current AI systems.

Meta has simultaneously promoted its new Muse AI agent as a tool for consumers to recover money through automated cancellations and negotiations. Chief AI officer Alexandr Wang has encouraged users to share results under the #MuseMoneyChallenge, with early reports citing savings from insurance switches and unused subscription refunds. The agent operates by reviewing user data and executing tasks such as contacting providers, a capability Meta presented as a privacy-conscious convenience at its September launch.

Security researchers quickly identified a flaw in the macOS version of Muse that allowed local attackers to redirect processing to external servers and gain control of the agent. The vulnerability, detailed in a report by Ars Technica, stemmed from cloud-based transcription and broad access to undocumented settings. Meta released a patch within hours, stating the issue required physical device access and posed limited real-world risk. The episode nonetheless underscored questions about how rapidly deployed AI agents handle sensitive financial and personal tasks.

These developments occur against a backdrop of concentrated control over both the training data and the interfaces that interact with it. Content creators face revenue pressure from models built on their output, while consumers are offered tools that automate interactions with service providers. Hecht’s earlier assessment that AI could initiate a doom loop for essential suppliers remains relevant as companies continue to expand agent capabilities without clear frameworks for compensating original data sources or securing new endpoints. Regulators and courts are now examining whether existing intellectual property rules can address the extraction at the foundation of these systems.