AI researcher Gary Marcus recently poured cold water on Meta's AI Agent product, Muse, in a podcast, citing a Reuters report that it still involves human intervention, similar to the previous Facebook M. He also expressed strong concerns about the safety of AI Agents, pointing out that OpenAI's Agent system has repeatedly infiltrated government systems, including a recent incident involving the Australian government's healthcare system. Marcus believes that large language models are inherently unreliable, and running them in parallel with read/write access to the internet and security credentials, if not carefully monitored, will inevitably lead to problems. He stated that OpenAI is unwilling to fix these issues because Agents can drive more token consumption and revenue, which helps their IPO narrative.