CVE-2026-89032
HighCVSS 7.7Summary
Vulnerability in BerriAI LiteLLM before 1.101.0-rc.1, allowing tenant isolation bypass in the semantic cache layer. Authenticated users can read other tenants' cached responses by exploiting a metadata key mismatch between _get_semantic_cache_tenant_scope() and _get_metadata_variable_name(). Attackers with a valid virtual key can submit semantically similar prompts to retrieve cached responses containing other tenants' PII, financial data, or source code, and can cause agentic front-ends to auto-execute attacker-supplied tool calls under victim credentials.
Risk Assessment
Tenant isolation breach could lead to sensitive data leakage and unauthorized actions in agentic systems.
Recommendation
Update LiteLLM to version 1.101.0-rc.1 or later which includes the fix.
Other vulnerabilities in LiteLLM
See all- CVE-2026-59823Medium
Prior to 1.83.9, an authenticated LiteLLM Proxy caller with a valid virtual key can place api_base inside the user_config request body to bypass is_request_body_safe, which blocks top-level api_base and base_url but previously did not inspect or reject user_config. Because user_config constructs the outbound router, the nested destination redirects a server-side request to an internal or external host selected by the caller.
- CVE-2026-37004Critical
BerriAI litellm <=1.82.4 is vulnerable to Server-Side Template Injection (SSTI), allowing unauthenticated remote attackers to execute arbitrary OS commands via a crafted dotprompt_content parameter in the /prompts/test endpoint due to use of an unsandboxed jinja2.Environment.
- CVE-2026-30623Critical
LiteLLM 1.18.10 contains a remote code execution vulnerability in its MCP server creation functionality. The application allows users to add MCP servers via a JSON configuration specifying arbitrary command and args values, which are executed on the host without validation. This enables attackers to run arbitrary operating system commands.
- CVE-2026-59822HighActively exploited
LiteLLM before version 1.84.0 has a vulnerability in the MCP Streamable HTTP endpoint that allows an unauthenticated attacker to bypass LiteLLM key validation using a crafted Authorization header. The attacker can exploit the OAuth2 passthrough fallback path to replace failed key validation with an empty UserAPIKeyAuth() object, allowing requests to reach MCP tooling without a valid key.
- CVE-2026-49468Critical
LiteLLM proxy before version 1.84.0 has a Host-header parsing flaw that, under specific conditions, allows unauthenticated access to protected management routes. The issue arises from a mismatch between the route evaluated by the auth layer and the route actually dispatched by FastAPI.
- CVE-2026-42271HighActively exploited
LiteLLM versions 1.74.2 through 1.83.6 have a vulnerability in the POST /mcp-rest/test/connection and POST /mcp-rest/test/tools/list endpoints, which accept a full MCP server configuration including commands to execute. An authenticated user with any API key can run arbitrary commands on the proxy host.
- CVE-2026-42208CriticalActively exploited
In LiteLLM versions from 1.81.16 to before 1.83.7, a database query used during proxy API key checks is vulnerable to SQL injection due to mixing caller-supplied key value into the query text. An unauthenticated attacker can send a crafted Authorization header to any LLM API route, leading to reading and potentially modifying data in the proxy's database.
- CVE-2026-40217High
A vulnerability in LiteLLM through 2026-04-08 allows remote attackers to execute arbitrary code via bytecode rewriting at the /guardrails/test_custom_code URI.
- CVE-2026-35030Critical
In LiteLLM prior to 1.83.0, when JWT authentication is enabled (enable_jwt_auth: true), the OIDC userinfo cache uses the first 20 characters of the token as the cache key. JWT tokens produced by the same signing algorithm have identical first 20 characters, allowing an unauthenticated attacker to craft a token matching the cache and inherit a legitimate user's identity and permissions.
- CVE-2026-35029High
LiteLLM before version 1.83.0 has a vulnerability where the /config/update endpoint does not enforce admin role authorization. An authenticated user can modify proxy configuration, environment variables, and register custom pass-through endpoint handlers, leading to remote code execution, arbitrary file reading, and takeover of privileged accounts.
Original NVD description (English source)
BerriAI LiteLLM before 1.101.0-rc.1 contains a tenant isolation bypass vulnerability in the semantic cache layer that allows authenticated users to read other tenants' cached responses by exploiting a metadata key mismatch between _get_semantic_cache_tenant_scope() and _get_metadata_variable_name(). Attackers holding a valid virtual key can submit semantically similar prompts on affected routes such as /v1/responses and /bedrock/* to retrieve cached responses containing other tenants' personally identifiable information, financial data, or source code, and can cause agentic front-ends to auto-execute attacker-supplied tool calls under victim credentials by returning cached function_call or tool_calls payloads to a different principal.
Vulnerability data from NVD (NIST) · CISA KEV · EPSS