How the attack works
An attacker with a normal internal_user account calls one of LiteLLM's key-management endpoints (/key/generate, /key/update, /key/regenerate, or /key/service-account/generate). In the request they set the allowed_routes parameter to include administrative or management routes — such as user, key, team, or global management endpoints — or use a wildcard to cover all routes. LiteLLM does not validate this list against the caller's own role, so it issues a key that carries those elevated permissions. The attacker then uses that key to call admin-only functionality, effectively becoming a proxy_admin.
Why it matters
An organization running an affected LiteLLM proxy can have any authenticated low-privilege user escalate to full administrative control over the proxy, including managing other users, teams, and keys — without any additional exploit step.
What you can do
- Upgrade LiteLLM to 1.83.14 or later.
- Audit existing API keys for allowed_routes values that include administrative or wildcard routes, especially those created or updated by non-admin accounts.
- Restrict which accounts can call /key/generate, /key/update, /key/regenerate, and /key/service-account/generate until patched.
- Monitor key-management requests for allowed_routes values reaching admin/management endpoints and confirm each one was created intentionally by a proxy_admin.
Known benign look-alikes
- A proxy_admin generating a key with broad allowed_routes intentionally — expected to match; review confirms intent.
- Security advisory or PR text quoting the CVE-2026-47101 payload.
- Key generation scoped to inference routes only (/chat/completions, /embeddings, /v1/models).