UCanTransit for Copilot
The AI review rates the findings as likely false positive, but the risk score (53/100) still counts them.
Analysis record
- Analysed
- 3 weeks ago 25 October 2026
- Version
- v0.6.0
- Artifact
- SHA256 b3da348b…b4b90f
- Detail
- Full list
Is UCanTransit for Copilot safe?
ucantransit lets you point GitHub Copilot Chat at a model endpoint you supply. You paste a URL and an API key, and it works out whether the far end speaks the OpenAI, Anthropic or Gemini protocol. It declares no special permissions and no host permissions, so it runs on what VS Code already gives a chat model provider. The endpoints in the code are the ones this job needs: openrouter.ai, wafer.ai, inference.net and models.dev. Two fetch calls, in extension/out/http.js and extension/out/modelCatalog.js, carry your prompts out and the model list back.
Much of the scanner output is strings that look like hostnames and are not. Entries such as choice.delta, text.delta, entry.vision and record.vision are field reads on streaming API responses, the same names every OpenAI-compatible client touches. Taking .delta off a chunk is not a network destination. The README badge URL img.shields.io and the document schema URL schemas.openxmlformats.org land in the same list for the same reason.
Nothing here opens your .env, your SSH keys or VS Code secret storage. The one key involved is the one you type in yourself. No install-time script runs and no malware signature matched, so the scanner tripped on property names and on build output under extension/out, which is compiled JavaScript rather than hidden code.
The thing worth knowing is design, not deception. Whatever you type into Copilot Chat reaches whichever endpoint you configure, and that endpoint sees it. Point it at a service you trust.
Evidence ledger
Ranked by severity · findings with a source location link to the code viewer
YARA Rule Matches
8 rules| Severity | Rule | Hits | Files | Metadata |
|---|---|---|---|---|
| LOW | credential env files | 1 | out/provider.js | - |
| LOW | postinstall persistence mechanism | 1 | out/displayName.js | - |
| LOW | postinstall file manipulation | 9 | out/modelCatalog.jsout/protocols/responses.jsmedia/panel.js +6 more | - |
| LOW | postinstall system command | 9 | out/protocols/openai.jsout/protocols/types.jsout/protocols/gemini.js +6 more | - |
| LOW | postinstall obfuscation | 7 | out/reasoningCatalog.jsout/protocols/anthropic.jsout/modelsDev.js +4 more | - |
| LOW | postinstall network communication | 12 | out/provider.jsLICENSE.txtout/panel.js +9 more | - |
| LOW | postinstall file download | 8 | out/reasoningCatalog.jsout/modelCatalog.jsreadme.md +5 more | - |
| LOW | NoUseWeakRandom | 5 | out/protocols/responses.jsout/protocols/anthropic.jsout/panel.js +2 more | - |
Publisher Evidence
LowZoneYottaZenith
Publisher identity, store signals, distribution reach, and warning signals used for context. Treat this as supporting evidence, not a clean bill of health.
12 evidence rows available.
Finding Categories
YARA Rules Matched
8 rules(52 hits)AI Security Report
AI Security Review
Evidence context: threat category none; evidence quality moderate.
What the extension is for
ucantransit 0.6.0 on OpenVSX, 126 installs, lets a user paste an endpoint URL and an API key and then route GitHub Copilot Chat through that endpoint. It reads the OpenAI, Anthropic and Gemini protocol shapes and can run several sites at once. That purpose sets the bar for what access is reasonable: HTTP requests out, replies parsed back.
Filesystem and process access against that purpose
The manifest declares no permissions and no host permissions. No postinstall finding exists. No process-spawn or shell indicator appears anywhere, and there is no critical or high severity entry in the results. The two capability findings are NET-FETCH-extension/out/http.js-36, which performs the chat request, and NET-FETCH-extension/out/modelCatalog.js-116, which retrieves a model list. An extension whose entire function is talking to a remote model API must make HTTP calls; these two match the feature set one for one. There is no workspace enumeration, no file walk, no write outside the extension's own config path.
Credential access
There is nothing in the secret category. No finding touches .env, .git/config, .ssh or VS Code secret storage. The only credential in play is the API key the user types into configuration, sent to the URL the user also typed. That is the transaction the extension advertises.
The strongest counterargument
A router that accepts an arbitrary base URL is a courier for anything you paste into chat, and a hostile endpoint would see your prompts. That exposure comes from the endpoint the user chooses, and it is the same exposure every OpenAI-compatible client carries. extension/out/http.js shows prompt data going to a configured base URL, and nothing in the results shows a second, hidden destination.
The second counterargument is the endpoint list itself. It looks alarming until you read it. choice.delta, text.delta, entry.vision, model.vision, guess.vision and record.vision are property chains lifted off streaming response objects, the standard fields of an OpenAI-style chunk, not hostnames. The domains that are real fit the purpose: openrouter.ai, wafer.ai, inference.net and models.dev are model providers or catalog sources. img.shields.io and schemas.openxmlformats.org are README badge and document-schema URLs.
The remaining volume is 52 low-severity code smells. The postinstall_* and credential_* YARA rules fire on ordinary Node patterns such as fetch, child_process imports and process.env reads, and they fire in bulk on bundled output under extension/out. Finding count tracks bundle size, not intent.
What would change this
A fetch to a domain that is not user-configurable, a read of workspace files bundled into the request body, or an activation-time download of remote code. None of those appear in http.js, modelCatalog.js or anywhere else in the results.
Key Reasons
- Only capability findings are two network fetches, NET-FETCH-extension/out/http.js-36 and NET-FETCH-extension/out/modelCatalog.js-116, both required by the stated purpose of routing chat to a user-supplied model API.
- Manifest declares no permissions and no host permissions; no postinstall, process-spawn or critical/high severity finding appears.
- Zero findings in the secret category; the only credential handled is the user's own API key sent to the user's own configured URL.
- Real domains in the endpoint list (openrouter.ai, wafer.ai, inference.net, models.dev) are model providers and catalog sources; the rest of the IoC list is property chains like choice.delta and text.delta.
- All 52 code-smell entries are low severity and match bundled Node patterns that postinstall_* and credential_* YARA rules fire on by design.
False Positive Considerations
- IoC extractor reading streaming-response property chains (choice.delta, text.delta, entry.vision, record.vision) as domains
- README badge and schema URLs (img.shields.io, schemas.openxmlformats.org) counted as network endpoints
- Low-severity code-smell rules (postinstall_*, credential_*) firing on bundled compiled output in extension/out
- Legitimate model-provider domains (openrouter.ai, wafer.ai, inference.net, models.dev) inflating the IoC count
Reviewed 2026-10-28 · version 0.6.0; recommended action: no action; model confidence 75%.
Source Code Not Available
Source code is not available for this version of the extension.
About This Extension
Frequently Asked Questions
Similar Extensions
Related extensions from the same publisher or marketplace