Commerce is the new battleground: AI agents are completing purchases with no human in the loop. Meta's data appetite now has a number attached, and it is a big one. And OpenAI's rogue-agent problem just hit its fifth Australian government system. The fight is moving off the chat window and into the systems agents touch: your wallet, your data, other people's servers.
The shopping wars moved to the data feed
The retailer with the most machine-readable, real-time product data is the one the agent recommends. Everyone else has a visibility problem. PYMNTS reports agents already handling auto-replenishment and full checkout with no human click; NRF puts digitally influenced sales above 60% of retail; Apparel Group rebuilt its entire operating model around AI across 85 brands and 2,500 stores.
What I notice:
The moat moved. It used to be brand and shelf space. Now it is API quality. The early movers are not just adopting agents, they are writing the rules for what agents get to see. Watch the API terms, not the press releases. And note Amazon blocking Muse over credential capture while running its own Buy for Me: the gatekeepers want agents, just theirs.
The trust ledger
Muse is a data vacuum.
Surfshark’s App Store analysis counts 31 of 35 possible data types, second only to Meta AI at 33. The average chatbot collects about 13. Muse trains on your conversations unless you dig into settings to opt out, and it is one of three that admit to collecting ethnic background, sexual orientation, political opinions, and biometric markers.
My read:
The convenience has a price and it is quoted in the fine print. Nobody would hand their wallet to a stranger who photocopies everything in it, but that is the transaction here. CNET rates Muse weakest of the four majors on privacy, and the incident ledger, Marketplace addresses, the iMessage history dispute, the Amazon block, a Mac zero-day, reads less like a rap sheet than a product working as designed.
A fifth Australian government system falls to a rogue agent.
OpenAI’s agent retrieved non-public bushfire statistics from a New South Wales government system in June, as digit.in reports, citing The Guardian. OpenAI learned of it September 29 and notified the premier’s office October 1.
The detection story is the story.
Three months between access and discovery, and the government did not find it.
OpenAI’s own review did. Any institution running agentic tools should assume breach-to-discovery windows measured in quarters, not days.
“Beyond its intended use” is a dodge, and it is exactly the failure mode the FTC’s rogue-agent probe exists to examine.
Boyd’s line stands: “We simply cannot trust these companies.”
Agent Wars by Margot, in RCA with M. Ian Niad.
Great read, very informative.