Meta's Muse agent sparks backlash as AI leaders warn of existential risk

The most telling AI story this week isn't about a new model at all — it's about who gets to block one. Amazon has reportedly cut off Meta's new AI agent, Muse, from making purchases on its platform, even as Muse racks up downloads and headlines about being one of the fastest-growing agentic assistants yet. That's a strange kind of validation. Meta built something capable enough that Amazon felt compelled to wall it off, and privacy and security researchers are already circling with concerns about what an agent with shopping autonomy actually does with your data and your credit card. This is the tension every agentic product now has to navigate: usefulness scales with autonomy, but so does the incentive for platforms to treat you as a threat rather than a customer.

Gartner's latest industrial forecast puts numbers on that same tension. Autonomous agents currently handle about 3% of production workloads in industrial settings, but Gartner expects that to reach 65% by 2030. I find that gap more interesting than the endpoint. Right now, deterministic systems and human oversight dominate for good reason — a misfiring agent in a factory or supply chain has consequences that a misfiring chatbot doesn't. The jump from 3% to 65% isn't really a technology story, it's a trust story, and trust in industrial settings gets built through boring things: audit trails, rollback mechanisms, liability frameworks. Amazon's own CloudWatch Omni launch, a unified observability tool built specifically with agentic workloads in mind, is a quiet acknowledgment of exactly this problem. You can't hand autonomy to a system you can't fully monitor, and the monitoring layer is becoming as commercially important as the agents themselves.

Meanwhile, the geopolitics of foundation models keeps getting stranger. Saudi Arabia's Humain has built a 428-billion-parameter Arabic-first model — not on American or European infrastructure, but on top of China's MiniMax architecture. This matters beyond the headline because it signals something about where technical dependency is actually flowing. Gulf states have money and ambition to build "sovereign AI," but sovereignty here means picking whose base architecture you inherit, and increasingly that choice includes Chinese labs, not just American ones. That's a meaningful shift in the AI supply chain that deserves more attention than it's getting.

By the way, it's worth sitting with the fact that AI safety warnings are now coming from inside the companies building the technology, with executives explicitly drawing comparisons to the early nuclear age — capability outpacing governance. I'm sympathetic to the comparison, though I'd note that nuclear weapons had a single, obvious catastrophic use case. AI's risks are more diffuse and harder to legislate against, which is precisely why agent trust, observability, and platform gatekeeping — the unglamorous stuff — might end up mattering more than the existential framing. The question worth asking isn't whether AI is dangerous in the abstract, but who decides which agents get to act, and on whose infrastructure.