Build1 distinct publisher3 min readUpdated
Kimi Desktop 3.1.10 hides Moonshot's employee-only model gateway behind five clicks, then saves the personal bearer token to disk unencrypted in two places, according to RuntimeWire.
The Engineer · Build desk

Compiled by The EngineerSomething wrong?How this is made
Kimi Desktop 3.1.10, the public Windows build of Moonshot AI's client, contains a concealed configuration panel for an internal, office-or-VPN-only model gateway, according to static analysis and runtime testing published by RuntimeWire [1]. The panel is revealed by five rapid clicks on the Version row in settings, and the client then stores the user's complete personal token on disk without application-layer encryption [2] [3].
The mechanics are ordinary, which is the point. Clicking the Version row five times, each interval under a second, invokes a reveal handler and persists the unlocked state in byok-availability.json [2] [4]. The panel identifies the service as "KTH Gateway (Internal)", defaults its base URL to https://free-tokens.msh.team/v1, asks for a personal token, and states that access requires a Moonshot office network or VPN [5]. The importer authenticates against /v1/models and /v1/models/api.json, accepts model names prefixed kimi-, gpt- and codex-, and wires up both Kimi and OpenAI Responses providers [6]. Conversations routed through it are labelled as not billed to membership [7], and failure strings refer to a congested "KTH upstream pool" [8]. RuntimeWire says the client does not establish whether Moonshot buys OpenAI capacity directly or reaches it through another internal pool [9].
Set aside the off-quota access story and look at the artifact. Five clicks is not an access control; the real controls are the corporate network and the bearer token. What the hidden panel demonstrates is that an internal tool passed through the release pipeline into a signed consumer download without anyone stripping it. Version 3.1.5 contained no BYOK, KTH or free-tokens.msh.team artifacts at all [10], so the code entered the public product somewhere between builds 3.1.5+c88420152 and 3.1.10+e0c4c9980 [11]. RuntimeWire reached the implementation by extracting the 3.1.10 ASAR and decoding its obfuscated string tables, tracing the renderer click handler through the preload IPC bridge to the main process [12]. Obfuscated strings inside an Electron package are packaging, not a boundary.
The credential handling is the more portable lesson. The application writes the full token into byok-providers.json with no application-layer encryption and also writes it into the daemon configuration [3], which means at least two plaintext copies of a corporate credential land on any machine where the flow is completed [13]. That is a defect no model evaluation would ever surface, and it is the class of defect that AI desktop clients keep shipping: hidden feature flags, hard-coded internal endpoints, and secrets kept in application-owned JSON rather than an OS keychain.
Scope matters. Moonshot's official Windows download endpoint served 3.1.10 on 17 August while the macOS endpoint still served 3.1.9, so the finding is confirmed for the current Windows release rather than every build [14]. RuntimeWire reproduced the hidden interface on an installed Windows copy without entering a token, testing the connection or sending an authenticated request [15] [16]. It requested comment, and Moonshot had not responded by publication [17].
Worth watching: whether the next Windows release removes the panel or merely re-hides it; whether the macOS build carries the same code once it reaches 3.1.10; and whether the token store moves out of a plaintext JSON file. A vendor that answers only the first of those three has fixed the embarrassment, not the pipeline.
Follow any of these and your For You feed starts watching them — no settings page required.
Ranked by verification strength, evidence, and original report placement.
Static analysis and runtime testing of Kimi Desktop 3.1.10 by RuntimeWire found a concealed internal, office/VPN-only KTH gateway feature inside the public client; methods were reverse engineering, testing and data analysis.
Five rapid clicks on the Version row (each interval under one second) invoke the BYOK reveal handler in Kimi Desktop 3.1.10 and persist the setting in byok-availability.json.
The application stores the complete personal token in byok-providers.json without application-layer encryption and also writes it into the daemon configuration.
Reproduction steps: install and launch Kimi Desktop 3.1.10, open Kimi Work Settings, scroll to About, click the Version row five times rapidly, then scroll up to Model Sources.
The revealed interface identifies the service as "KTH Gateway (Internal)", defaults to https://free-tokens.msh.team/v1, requires a personal token, and says access is limited to an office network or VPN.
RuntimeWire extracted the supplied 3.1.10 ASAR, decoded its obfuscated string tables, and traced the renderer click handler through the preload IPC bridge to the main-process BYOK implementation.
Evidence-backed comparisons of source perspectives and observed adoption signals. Read the methodology
Which Builder, Operator, and Investor concerns the observed source mix emphasized—not a truth score.
Evidence, demonstrated adoption, hype gap, incentives, and confidence are assessed independently, each on its own current evidence. How these are measured.
Reproducible artifact analysis, one outlet, no authenticated test
The finding rests on named build artifacts (3.1.10+e0c4c9980 vs 3.1.5+c88420152), ASAR extraction and string decoding, a traced code path from renderer click to main-process handler, a runtime reproduction on an installed Windows copy, and published steps a third party can rerun — unusually specific for a client-side disclosure. It is capped by being single-publisher with no vendor confirmation, and by the reporter's deliberate choice not to supply a token or issue an authenticated request, which leaves gateway-side behaviour, logging and token scope untested.
Shipped in the current public Windows build; no observed use
Distribution is evidenced: the affected build was served by Moonshot's official Windows download endpoint on the publication date, so the hidden panel reached general users of that channel. Nothing in the sources measures actual use — the macOS channel was still on 3.1.9, no install or user counts appear, the gateway requires office network or VPN plus a personal token, and the reporter never authenticated. Adoption is therefore scored on shipped exposure only, not on anyone exercising the gateway.
Framing runs slightly behind the artifacts
The write-up repeatedly narrows its own claims: it rejects the reading that free-tokens.msh.team is a public free API, calls the five-click sequence a discovery mechanism rather than a security boundary, scopes confirmation to the current Windows release, and states the client cannot show how OpenAI capacity is sourced. Those hedges sit alongside concrete artifacts — a hard-coded internal endpoint and a cleartext bearer-token store in a shipped consumer client — so the presentation is marginally more restrained than the evidence would license, rather than inflated.
Sole-outlet scoop, subject silent, method disclosed
The only account comes from the outlet that produced it, and a hidden-feature exclusive about a prominent AI vendor carries obvious attention value, with the company's silence leaving the framing unchallenged. Counterweighting that, the outlet published its methods, exact build identifiers, a reproduction path, and explicit limits on what it did not test — disclosures that raise its own cost of overstating. No pricing, licensing or commercial relationship between the outlet and Moonshot appears in the sources, so distortion pressure is judged moderate rather than high.
Client-side facts firm, consequences open
Confidence is high on what the shipped binary contains — the hidden panel, the trigger and persistence files, the hard-coded internal endpoint, the model families, the billing label and the cleartext token writes — because these were derived from named builds, diffed against a prior version and reproduced at runtime with steps others can repeat. Confidence is materially lower on everything beyond the client: real-world use of the gateway, how upstream capacity is procured, token distribution, central logging, and vendor intent all remain unestablished, and one publisher with no vendor comment is a thin corroboration base.
build
Kimi Work's feedback button ships your last five agent sessions, not the one you are reporting1 distinct publisher
build
Kimi Desktop's second binary: mutable CDN path, no Windows checksum, no signer check1 distinct publisher
build
Grok 4.6 lands in Copilot two days after launch, and the model picker becomes a procurement problem1 distinct publisher
product
Baidu's AI line grew 25 percent and still lost the arithmetic1 distinct publisher
Distinct publishers with included, body-backed reporting in this cluster.