Meta tells Ray-Ban Display developers its untrusted-content flag is not enforced
Meta has a flag that web apps on Ray-Ban Display use to tell Meta AI that a piece of text came from somewhere untrustworthy, and Meta's own documentation tells developers not to rely on it. "Set untrustedContentHint honestly, since it's not an enforced security boundary", Meta's WebMCP agent tools page says. "Your tool's result text goes into the assistant's context, so returning unmarked untrusted content is how a prompt injection reaches the model."

The Meta Neural Band, the wristband wearers use to drive the glasses by hand when they are not speaking to the assistant. Image: Mixed News media library
The page is Meta's guide for developers building web apps that the assistant on the glasses can operate by voice, read at 22:04 UTC on October 7. Meta says the capability itself is "off by default and enabled per device", reaching glasses either through Developer Mode or "because the wearer is in the rollout", and that "WebMCP for Web Apps is coming soon". So this is guidance written ahead of a wide release, not an account of anything that has happened to a wearer.
Meta tells developers to distrust their own assistant
Meta tells developers not to trust the arguments the assistant hands them either: "Meta AI chooses the arguments. Nothing guarantees they match your inputSchema", it says, so developers should "treat input as untrusted and validate it in execute, exactly as you would a request body from the network". Then a plain security rule: "Never take a destructive or irreversible action solely because a tool was called with plausible-looking arguments."
Meta extends the same scepticism to what the assistant reports back. One design rule is simply "The page is the referee": "Never accept Meta AI's account of what happened. Tools return facts your page can verify. If an outcome can be scored, your page scores it." Where the model has to report what the wearer said and then act on it, Meta tells developers to "fuse both into a single call so there is no path to a result that skips the check".
Meta also tells developers not to put a real answer into a parameter description, because "a model will send the example", and to "use placeholder shapes, not plausible values". Every one of those rules points the same way, and all of them are Meta's: assume the text handed to the model can carry instructions, assume the arguments may not match the schema, and do not accept the model's account of what it did.
Seven names fail without telling you
The reserved tool names are a second trap. Seven of them are "reserved for the browser's own tools and are rejected": openUrl, goBack, goForward, reload, getCurrentUrl, getPageTitle and getPageText. Those are exactly the names a developer would reach for when wiring up navigation.
The failure is invisible. "The rejection happens after registerTool() has already resolved, so this failure is silent from the page's point of view", Meta writes. "The promise fulfills and the tool simply never reaches Meta AI." Meta puts the same warning more generally: "Native/provider rejection is not acknowledged back to the page, so these do not prove the assistant can see the tool."
All of this lands on glasses rather than a phone because of screen size. Meta gives the Ray-Ban Display panel as "600x600px" with no touchscreen, so "a task that takes five focus moves and three pinches through your UI can often be one spoken sentence instead". Speech is the cheap interaction at that size, which is also why the assistant ends up reading text from third-party apps at all.
Nothing here describes an attack, a disclosure or a patch. It is a set of instructions for developers of a capability most owners cannot switch on, on glasses that Meta priced at £749 in the UK and €899 in Germany. The notable part is who wrote it: the company shipping both the assistant and the glasses, telling developers in its own documentation where the model can be talked into something.