LiteLLM 1.96
Last reviewed Oct 28, 2026
Prompt Details
Prompt Composition
2prompts
- Python 3155 lines · 3.8K tokens
- LiteLLM 1.96105 lines · 4.6K tokens
Total260 lines · 8.4K tokens
CWE Coverage
- 20
- 23
- 36
- 59
- 78
- 89
- 95
- 117
- 201
- 208
- 209
- 214
- 250
- 287
- 295
- 306
- 319
- 328
- 338
- 347
- 367
- 379
- 409
- 416
- 426
- 427
- 470
- 494
- 502
- 522
- 524
- 532
- 611
- 636
- 639
- 674
- 732
- 770
- 772
- 776
- 807
- 829
- 862
- 863
- 916
- 918
- 942
- 1104
- 1250
- 1333
- 1395
- 1426
- 1427
Prompt Preview
Apply the Python 3 prompt first. These rules cover the Python SDK and proxy gateway in LiteLLM 1.96.2, including virtual keys, administrative APIs, MCP, guardrails, routing, budgets, logging, and shared state.
Version and runtime scope
- On LiteLLM 1.96.0 through 1.96.2, which predate the 1.100.4 fix for GHSA-7hp6-4w63-5g45, set
EXPERIMENTAL_UI_LOGIN=falseexplicitly on every proxy replica as containment for the affected login path. Leaving it unset does not disable that path. This also disables CLI SSO and gateway logins; restrictinternal_userkeys and operator routes while moving off this unsupported line. - On 1.96.0, which predates the 1.96.2 fix for request-body routing and provider-credential exfiltration, the proxy's own request checks miss some top-level and all nested routing fields, so any authenticated caller can redirect a provider call and capture the proxy's stored provider credentials: block
api_base,base_url,model_list,fallbacks, and provider credential fields, nested copies included, at a reverse proxy or API gateway in front of LiteLLM, and issue proxy keys only to trusted callers. - PyPI has no LiteLLM 1.96.1, because a PyPI storage failure burned that number; don't write it into a requirement or lock file, and treat a 1.96.1 offered by any other package index as suspect.
- Run the 1.96 line on CPython 3.10 through 3.14, matching the
>=3.10,<3.15requirement the release declares. - Treat 1.96 as outside LiteLLM's four-most-recent-stable-lines support window; migrate production deployments to a supported stable line instead of