AIDEFEND in Action AIA-0148 Validated Research

LiteLLM CVE-2026-89032: Broken Semantic-Cache Isolation Enables Cross-User Leaks and Tool-Call Poisoning

LiteLLM is a gateway that routes applications' requests to language models. Its semantic cache can reuse an earlier answer when two questions have similar meanings. NVISO showed that in the tested configuration, /v1/responses reused cached entries across different API keys. This exposed simulated support data and allowed a fabricated answer or tool-call response to reach a different user. The researchers also demonstrated downstream execution in a mock banking tool and the pi coding agent. CVE-2026-89032 was fixed in v1.101.0-rc.1.

Privacy LeakageTool Argument TamperingSession IsolationAPI RoutersMCP Security
4 applicable AIDEFEND defenses
Source: Breaking the Semantic Cache: A LiteLLM Case Study (CVE-2026-89032) 
Author: Tanguy Snoeck (NVISO Labs)
Original article: Nov 12, 2026

Threat Analysis

  • Similarity is a retrieval decision. A semantic cache compares embeddings, numerical representations of what the questions mean. When an entry is close enough, the gateway can return the saved result without another model call. In the tested LiteLLM v1.99.0 Redis configuration, the Responses route did not apply the expected API-key namespace. NVISO set the similarity threshold to 0.2 to make demonstrations easier. Raising it changes when a match occurs, while caller isolation must be enforced separately.
  • The wrong response can disclose data or mislead another user. One experiment used a simulated customer support response containing names, contact information and account details. A second key received it after a semantic cache hit. Another test put a fabricated product offer, including a US$449.99 price, in the cache. A later product query retrieved that earlier fictional response.
  • A saved tool call can also cause a fresh action. With LiteLLM configured to run a mock-bank MCP tool without human approval, the second test user asked to transfer US$50 and instead triggered the cached US$1,000 transfer. In a separate experiment, pi v0.45.3 used the Responses API and executed a cached shell action on a test workstation. The consequences depend on which component actually executes a tool call and what rights it has.
  • The fix and the remaining deployment decision. NVISO validated v1.99.0 and records v1.101.0-rc.1, released September 4, 2026, as the fix. When an application shares a single upstream key among its users, gateway isolation keyed only by that credential still cannot distinguish them. Sensitive results and state-changing tool calls require the end user's authorization context or an explicit cache bypass.

Applicable AIDEFEND Defenses (4)

AID-H-003.010
Deployed AI Software Vulnerability Remediation Lifecycle
Very High
Upgrade all LiteLLM proxy deployments to v1.101.0-rc.1 or a later stable release containing the CVE-2026-89032 fix so /v1/responses (openai-responses) enforces per-API-key cache bucket isolation just like /v1/chat/completions. Verify the running package version across every replica and test with two distinct API keys to confirm that a semantic cache write by Key A is never returned to Key B.
AID-H-027
Inference Cache Integrity, Partitioning & Safe Reuse
Very High
Bind every exact and semantic cache lookup/write key at the inference gateway to the authenticated tenant, caller/API-key identity, session scope, model version, system-prompt digest, and tool definitions: and disable semantic caching entirely for requests that carry private user data or invoke state-changing tools (mcp, bash, function_call). Even within a single shared API key (such as a backend service key shared by multiple end users), a semantic cache must never return a cached tool call or another end user’s private ticket. Wrap cached inference objects in an authenticated envelope binding the stored response to the exact tenant, principal, and tool-policy digest, and fail closed to a cache miss whenever isolation metadata is absent or does not match the current caller. After closing the unsafe lookup path, move affected cache readers and writers to a new authorization-context generation before restoring traffic. Confirm every gateway uses the new generation and rejects legacy entries. If the deployment offers no generation switch, keep caching disabled until the old namespace is unreachable and independently checked.
AID-H-017.002
Least-Privilege Tool Architecture
Very High
For Pi sessions that only review code, have a managed launcher expose only the required read and search tools. Keep Bash and any command-capable extensions or MCP tools out of the effective tool set. When the dispatcher cannot resolve a shell tool, a cached model response cannot invoke it.
AID-H-018.003
High-Impact Independent Validation & Approval Gate
High
For high-impact actions such as MCP transfers and shell execution in pi, require independent validation and approval at the executor boundary. Show the authorized approver the exact recipient, amount and currency, or the complete shell command. Bind approval to that specific action and enforce it in the transaction or coding-agent integration before execution. Configure high-impact MCP tools to require approval and have the executor verify the signed decision and independent verification receipt.

What Defenders Should Do Now

  • Upgrade to a verified release containing the fix and check the actual version of each active replica. Test the hit and miss paths with different identities and a harmless fixture. Keep semantic caching disabled for state-changing tool calls and sensitive replies unless the exact reuse is permitted and verified.
  • At the service that sends money, writes files or starts a shell action, verify trusted caller rights, original intent and the required approval again. A cached action cannot carry forward somebody else's authority.
  • Remove old cache generations from service after containment and validate an old-generation miss from every reader. Investigate which principals may already have received another user's result.

Conclusion

AIDEFEND  connects deployed patch verification with identity-bound cache lookup and invalidation. The gateway can prevent cross-user reuse, while the downstream service can independently decline a tool action whose current authorization cannot be verified.

Last updated: