Build1 publisher2 min readPublished
RuntimeWire finds a Trusted Network invite flow inside Meta's public Muse JavaScript
RuntimeWire found identifiers for Trusted Network invitations, connection reviews and safety codes in 69 public JavaScript files from Meta's Muse. The names show how two people would connect, but not what a connection may share, and that is the part agent builders need.
The Engineer · Build desk
Drafted by a language model from the sources cited here and checked against its claim ledger before publication. How we use AISend a correction
What happened
- Taken together, the names describe people choosing contacts, exchanging invitations, reviewing a connection and later removing it.
- The code also names a safety-code decision, retry and close action, with matching safety-code and safety-code-access screen views.
- RuntimeWire worked from a logged-out browser and did not bypass access controls or test a signed-in Muse account.
- Instinct founder Noah Shinn said on September 13 that his company's Trusted Person network, a working version of the same idea, was open to all Instinct users.
- Meta had not responded to RuntimeWire's request for comment by publication time.
Compiled by The EngineerSomething wrong?How this is made
Why it matters
- constraint Teams cannot scope an integration against Muse connections yet, because the permissions an accepted connection grants decide what data crosses between users, and those are unknown.
- exposure If Muse ships this, the safety-code step is the visible control over who joins a person's network, and how strong it is depends on a function RuntimeWire could not identify.
- precedent Instinct has set the bar at tasks completed across users, so a Muse launch will be judged on the jobs its connected agents actually finish.
The identifiers sit in four catalogs: user interactions, screen views, application surfaces and logging [4]. RuntimeWire calls them instrumentation declarations and cautions that such catalogs can retain experiments [5]. An entry such as `chat_trusted_network_connection_request_review_click` tells a reader that a review action exists inside chat and that tapping it gets counted [6]. It does not say what accepting the request grants, and RuntimeWire did not establish what connections allow [3].
The screen-view catalog lists a Trusted Network overview, views of individual people, invitation screens, an invite-link view and connection requests [7]. Contacts are managed from a settings surface named `settings_trusted_network_contacts`, and the logging catalog records chat connection requests [7]. I think that split is the most useful detail for builders. The contact list lives in settings, and approval happens in the conversation [6][7].
The safety-code screens are the hardest part to read. There is a retry action, so whoever designed the check expected it to fail sometimes [8]. RuntimeWire could not establish whether the code verifies cryptographic identities, requires a human to compare a code, or serves some other purpose [9]. "A screen name cannot answer those questions," RuntimeWire wrote [19].
Instinct's network is the working comparison. Shinn described assistants finding meeting times, coordinating dinner for 10 people and adjusting recurring tennis sessions when someone's calendar changed [16]. In each example, data crosses the link and the assistant acts on it. On the Muse side, RuntimeWire did not recover the screens' implementation or see two Muse agents communicate. It treats agent-to-agent coordination as an inference from the names [10].
The evidence package is built to be checked. RuntimeWire indexed 43 evidence locations with source URLs, SHA-256 hashes, exact offsets and the surrounding code [12]. The capture ran from 6:56 p.m. to 7:03 p.m. Chicago time on October 1, seven minutes in all [1][13][1]. It came to just under 4 MB [2]. Anyone holding the package can run `python3 tools/verify_evidence.py` against it [14].
What to watch
- A Meta announcement or help page defining what an accepted Trusted Network connection can read or do, for example whether it reaches a user's calendar.
- A signed-in look at the Muse safety-code screen showing whether it has people compare codes or checks identity keys.
- A fresh capture checked against RuntimeWire's 43 hashed locations, showing whether the identifiers persist, vanish as an experiment, or gain names for actions beyond invites and approvals.