CVE-2026-89032: LiteLLM: semantic cache tenant bypass leaks data
HIGHLiteLLM versions before 1.101.0-rc.1 have a tenant isolation flaw in the semantic cache: a metadata key mismatch lets any authenticated user retrieve other tenants' cached LLM responses by sending semantically similar prompts. Cached content can include PII, financial data and source code, and returned cached function_call or tool_calls payloads can make agentic front-ends auto-execute attacker-supplied tool calls under a victim's credentials, which raises this from a leak to potential action-on-behalf. Exploitation likelihood is currently modest (EPSS 0.27%, top 82th percentile, no public exploit, no Nuclei template, not in CISA KEV, SSVC: TRACK), but the attacker only needs a valid virtual key, which is routine in shared multi-tenant gateways. Upgrade to 1.101.0-rc.1 or later, and until then disable semantic caching on multi-tenant deployments and review cache hits across tenants.
What is the risk?
High (CVSS 7.7, scope changed). Attack is network-based with low privileges, but requires high complexity (crafting semantically close prompts) and user interaction (a victim front-end consuming the poisoned or cross-tenant cache entry). Exploitation signals are low: EPSS 0.00272, no public exploit, not in KEV, SSVC TRACK. Confidentiality and integrity impact are high because cross-tenant data and tool calls are affected. Exposure is greatest for shared LiteLLM proxy gateways with semantic caching enabled and exposing /v1/responses or /bedrock/* routes. Confidence: MEDIUM (based on vendor commit and VulnCheck advisory; dependents count unknown).
How does the attack unfold?
What systems are affected?
| Package | Ecosystem | Vulnerable Range | Patched |
|---|---|---|---|
| LiteLLM | pip | — | No patch |
Do you use LiteLLM? You're affected.
How severe is it?
What is the attack surface?
What should I do?
1 step1) Upgrade LiteLLM to 1.101.0-rc.1 or later (see commit 16db51e and PR #39590). 2) Until patched, disable semantic caching (or caching altogether) on shared/multi-tenant proxies, especially for /v1/responses and /bedrock/* routes. 3) Flush existing semantic cache stores (Redis/vector backend) after patching to remove potentially cross-contaminated entries. 4) Do not cache responses containing tool calls/function_call payloads; require agent front-ends to validate tool calls against the requesting principal before execution. 5) Detection: audit proxy logs for cache hits where the requesting key/team differs from the key that created the entry, and for unusual bursts of near-duplicate prompts from a single virtual key. 6) Rotate or scope virtual keys and apply least-privilege on tools exposed to agents.
What does CISA's SSVC say?
Source: CISA Vulnrichment (SSVC v2.0). Decision based on the CISA Coordinator decision tree.
How is it classified?
Which compliance frameworks are affected?
This CVE is relevant to:
Frequently Asked Questions
What is CVE-2026-89032?
LiteLLM versions before 1.101.0-rc.1 have a tenant isolation flaw in the semantic cache: a metadata key mismatch lets any authenticated user retrieve other tenants' cached LLM responses by sending semantically similar prompts. Cached content can include PII, financial data and source code, and returned cached function_call or tool_calls payloads can make agentic front-ends auto-execute attacker-supplied tool calls under a victim's credentials, which raises this from a leak to potential action-on-behalf. Exploitation likelihood is currently modest (EPSS 0.27%, top 82th percentile, no public exploit, no Nuclei template, not in CISA KEV, SSVC: TRACK), but the attacker only needs a valid virtual key, which is routine in shared multi-tenant gateways. Upgrade to 1.101.0-rc.1 or later, and until then disable semantic caching on multi-tenant deployments and review cache hits across tenants.
Is CVE-2026-89032 actively exploited?
No confirmed active exploitation of CVE-2026-89032 has been reported, but organizations should still patch proactively.
How to fix CVE-2026-89032?
1) Upgrade LiteLLM to 1.101.0-rc.1 or later (see commit 16db51e and PR #39590). 2) Until patched, disable semantic caching (or caching altogether) on shared/multi-tenant proxies, especially for /v1/responses and /bedrock/* routes. 3) Flush existing semantic cache stores (Redis/vector backend) after patching to remove potentially cross-contaminated entries. 4) Do not cache responses containing tool calls/function_call payloads; require agent front-ends to validate tool calls against the requesting principal before execution. 5) Detection: audit proxy logs for cache hits where the requesting key/team differs from the key that created the entry, and for unusual bursts of near-duplicate prompts from a single virtual key. 6) Rotate or scope virtual keys and apply least-privilege on tools exposed to agents.
What systems are affected by CVE-2026-89032?
This vulnerability affects the following AI/ML architecture patterns: LLM gateways/proxies, multi-tenant LLM inference, agent frameworks, semantic caching layers.
What is the CVSS score for CVE-2026-89032?
CVE-2026-89032 has a CVSS v3.1 base score of 7.7 (HIGH). The EPSS exploitation probability is 0.27%.
What is the AI security impact?
Affected AI Architectures
MITRE ATLAS Techniques
AML.T0012Valid AccountsAML.T0040AI Model Inference API AccessAML.T0053AI Agent Tool InvocationAML.T0057LLM Data LeakageAML.T0085Data from AI ServicesCompliance Controls Affected
What are the technical details?
Original Advisory
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.
Exploitation Scenario
An attacker obtains a legitimate low-privilege virtual key on a shared LiteLLM gateway used by several business units. They submit prompts to /v1/responses that are semantically similar to likely queries from another tenant, such as requests to summarize a financial report or debug internal code. Because of the metadata key mismatch, the tenant scope is not enforced and the semantic cache returns the other tenant's cached answer containing sensitive data. In a second variant, the attacker seeds the cache with a response containing a malicious tool_calls payload; when a semantically similar prompt from a victim's agentic front-end hits the cache, the front-end auto-executes the attacker-supplied tool call using the victim's credentials.
Weaknesses (CWE)
CWE-863 — Incorrect Authorization: The product performs an authorization check when an actor attempts to access a resource or perform an action, but it does not correctly perform the check.
- [Architecture and Design] Divide the product into anonymous, normal, privileged, and administrative areas. Reduce the attack surface by carefully mapping roles with data and functionality. Use role-based access control (RBAC) [REF-229] to enforce the roles at the appropriate boundaries. Note that this approach may not protect against horizontal authorization, i.e., it will not protect a user from attacking others with the same role.
- [Architecture and Design] Ensure that access control checks are performed related to the business logic. These checks may be different than the access control checks that are applied to more generic resources such as files, connections, processes, memory, and database records. For example, a database may restrict access for medical records to a specific database user, but each record might only be intended to be accessible to the patient and the patient's doctor [REF-7].
Source: MITRE CWE corpus.
CVSS Vector
CVSS:3.1/AV:N/AC:H/PR:L/UI:R/S:C/C:H/I:H/A:NReferences
Timeline
Related Vulnerabilities
CVE-2026-1146129.9LiteLLM: forged admin token enables RCE via MCP stdio
Same package: litellmCVE-2026-370049.8LiteLLM: unauth SSTI in /prompts/test enables RCE
Same package: litellmCVE-2026-422089.8LiteLLM: SQL injection exposes LLM API credentials
Same package: litellmCVE-2026-306239.8LiteLLM: RCE via unsanitized MCP server config
Same package: litellmCVE-2026-543529.6Budibase: zip symlink bypass exposes all server secrets
Same package: litellm