Amazon blocks Meta's AI shopping agent, exposing a gap in online trust

Amazon has blocked Muse, Meta's AI shopping agent, from making purchases on its site. Amazon says Meta never asked for access, and that Muse does not announce itself as software while it browses. Amazon also says Muse appears to capture and store customer credentials. Meta disputes the last point.

What each side says

Meta says login details go into secure storage that the AI model itself never reads. Amazon's terms now require AI agents to identify themselves with a short line of text in their web requests. Muse does not send one. The facts of the dispute come from the two companies' claims, and the material here does not settle who is right.

The bigger gap

The article is an opinion piece by Sharat Menon, an engineering manager at Schneider Electric. His point is that this is an old security problem in new clothes. One company's automated software is acting inside another company's systems, and there is no shared way to confirm who or what is there.

For people, that was solved years ago. Standards called OIDC and SAML let you sign in to a partner's app without that partner handling your password. Nothing like that has been widely adopted for AI agents.

Work is under way. The OpenID Foundation has published a whitepaper and draft extensions such as OIDC-A. NIST has issued a concept paper. OWASP lists identity and privilege abuse as a top risk for agent-based apps. Menon says the missing piece is adoption.

Why a text label is not enough

Menon compares Amazon's rule to asking someone to state their name at an unmonitored door. A plain text label can be faked and proves nothing about who authorized the agent. Amazon also reportedly spotted Muse by its traffic patterns and blocked it. He calls that expensive and brittle.

What works today

Menon points to controls already used when AI coding tools act inside a single company. They hint at what a wider standard will need.

• Give the agent its own registered identity, linked to the person who authorized it.
• Check each action against a clear policy outside the agent, not the agent's own judgment.
• Let agents research and compare prices, but require a human to approve spending or account changes.
• Keep tamper-proof logs so a bad outcome can be traced afterward.

What to do

If you let AI tools act on your accounts or your customers' accounts, treat them as a separate user, not as you. Know who authorized each action, decide which steps always need a person to sign off, and keep a record of what the tool did.