Skip to content

chore: update litellm model catalog for v1.15.45 - #9589

Closed
Vidit Ostwal (Vidit-Ostwal) wants to merge 1 commit into
mainfrom
chore/update-litellm-catalog-v1.15.45
Closed

Vidit Ostwal (Vidit-Ostwal) wants to merge 1 commit into
mainfrom
chore/update-litellm-catalog-v1.15.45

Conversation

@Vidit-Ostwal

Copy link
Copy Markdown
Member

No description provided.

@strix-security

strix-security Bot commented Nov 20, 2026 •

Copy link
Copy Markdown

Strix Security Review

✅ No open security findings.

PR summary: Regenerates the vendored LiteLLM model catalog (`model_prices_and_context_window.json`): adds `successor_model` pointers to deprecated models, seven new model entries, and refreshed pricing/capability values. No code changes.

Review summary

This pull request only refreshes the bundled LiteLLM model catalog: it adds successor_model pointers to ~500 deprecated entries, introduces seven new model entries, and updates pricing, source, metadata and modality values. Reviewing every added and removed line, plus all runtime readers of the catalog, found no security issue: the bundled file is read only by the context-window resolver, which consumes max_input_tokens and litellm_provider as integer/identity values for message-compaction budgets, and the new fields have no reader in the codebase. The file parses cleanly, contains no duplicate keys, introduces no shadowing or changed limits for existing models, and a parse failure aborts import rather than falling back to anything permissive. Remaining effects are cost/truncation correctness rather than anything an attacker can turn into access, data crossing or code execution.

Ruled out (4)
  • Can the new `successor_model` and metadata fields redirect a request, leak data, or render text somewhere?
    Nothing in the repository reads `successor_model`, `metadata`, `supported_endpoints` or the new cost fields, and there is no deprecation-following or auto-substitution logic that could act on them.
  • Can an added or edited catalog value reach a dangerous sink such as a URL, api_base, provider routing, subprocess or log output?
    The packaged catalog is read only by the context-window resolver, which extracts `litellm_provider` and a type-checked integer `max_input_tokens` that feeds token budgeting alone.
  • Could the new entries shadow an existing model, break JSON parsing, or make a consumer fail open?
    The file parses cleanly with no duplicate keys anywhere, adds seven keys without removing or renaming any, changes zero existing context-window values, and parse failure raises at import rather than falling back to permissive defaults.
  • Do the added model IDs widen what a model name allowlist exposes or change what the CLI offers?
    The only allowlist-shaped use is telemetry name masking, which forwards a name only when it exactly equals a published catalog key and otherwise masks it, so new public model IDs cannot unmask a customer-specific name.

Review 1 on this PR · 2ef8a6f · requested Nov 20 04:19 UTC when the PR was opened


Reviewed by Strix
Re-run review · Configure security review settings · Try a deeper review with @strix-security /deep

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant