fix(passthrough): enforce caller model ACL on provider passthrough (#449) - #489
Conversation
Generic /passthrough/{provider}/* picked the first Model matching the
provider and injected its credentials for ANY valid API key, ignoring
the key's allowed_models. A low-privilege key could thus reach a
provider's upstream credentials (e.g. Jina/Cohere) it was never granted.
Require the authenticated key to be allowed to access a model of the
target provider: select the first provider model the key can access, and
reject with 403 when the provider has models but none are in the key's
ACL. Mirrors LiteLLM, which enforces the key's model access on
passthrough when a target is identifiable.
Routing target authorization (#29) is left consistent with LiteLLM's
model-group semantics: the requested virtual-router name is authorized
(already enforced); its operator-configured targets are not separately
re-authorized.
Fixes #449
|
Warning Review limit reached
More reviews will be available in 49 minutes and 1 second. Learn how PR review limits work. Your organization has run out of usage credits. Purchase more in the billing tab. ⌛ How to resolve this issue?After more reviews become available, a review can be triggered using the We recommend that you space out your commits to avoid hitting the rate limit. 🚦 How do rate limits work?CodeRabbit enforces hourly rate limits for each developer per organization. Our paid plans include higher PR review limits than trial, open-source, and free plans. In all cases, reviews become available again over time. During sustained high-volume PR review activity, CodeRabbit may temporarily slow when the next review becomes available. Please see our Fair Usage Limits Policy for further information. ℹ️ Review info⚙️ Run configurationConfiguration used: Organization UI Review profile: CHILL Plan: Pro Run ID: 📒 Files selected for processing (2)
Comment |
There was a problem hiding this comment.
Pull request overview
Fixes a security issue where /passthrough/{provider}/*rest would lend any provider's credentials to any valid API key, ignoring the key's allowed_models ACL. Now the gateway requires the authenticated key to be permitted for at least one model of the target provider; otherwise it returns 403.
Changes:
- In
dispatch()for passthrough, the model lookup additionally requiresauth.key().can_access(display_name). - Distinguishes "provider unknown" (404
ModelNotFound) from "provider known but ACL denied" (403ModelForbidden). - Adds an e2e regression test verifying the ALLOWED key gets 200 and the DENIED key gets 403.
Reviewed changes
Copilot reviewed 2 out of 2 changed files in this pull request and generated no comments.
| File | Description |
|---|---|
crates/aisix-proxy/src/passthrough.rs |
Enforces caller's model ACL when selecting a provider model for credential injection; returns 403 when no allowed model exists. |
tests/e2e/src/cases/passthrough-model-acl-e2e.test.ts |
New e2e test asserting ACL enforcement on generic passthrough. |
💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.
Problem
/passthrough/{provider}/*restselected the first Model matching the provider and injected that Model's provider credentials for any valid API key, ignoring the key'sallowed_models. A low-privilege key could therefore reach a provider's upstream credentials (e.g. Jina/Cohere rerank) it was never granted (#19/#20/#27).Fix
Require the authenticated key to be allowed to access a model of the target provider before lending credentials: select the first provider model the key
can_access, and return 403 when the provider has models but none are in the key's ACL. This mirrors LiteLLM, which enforces the key's model access on passthrough when a target is identifiable.On #29 (routing targets)
Left consistent with LiteLLM's model-group semantics: the requested virtual-router name is authorized (already enforced in
chat.rs), and its operator-configured targets are not separately re-authorized (LiteLLM authorizes the requested model-group name, not each underlying deployment).Behavior change
A key with no model of a given provider in its
allowed_modelsnow gets 403 from that provider's passthrough (previously 200, leaking credentials).Tests
E2E
passthrough-model-acl-e2e.test.ts: a key whose ACL excludes any openai model gets 403 on/passthrough/openai/*; a key allowed for an openai model gets 200. Existing passthrough e2e still passes.Fixes #449