Analysis
Where tenant boundaries break: two advisories
What a tenant is, how isolation fails in AI systems, and what the Spring AI and LiteLLM advisories ask of anyone building on shared infrastructure.
What a tenant is
A tenant is a customer, organisation or workspace with its own users, data and access rules inside a shared service. Two companies can use the same application while each should reach only its own records. Keeping them apart is tenant isolation.
AI products make this harder because data sits in more places. An assistant used by several companies may share one vector store, one memory service and one response cache, with each company's documents, conversations and saved answers separated only by logic in the application.
How it goes wrong
A tenant ID alone protects nothing. Isolation fails when the system doesn't apply it consistently: a database query omits the restriction, a search tool loses it while translating a request, or a cache returns a saved answer without checking whose it was. Both users logged in correctly, and one of them still receives the other's data.
Two published advisories show different ways this boundary can fail.
Spring AI: the filter that changes meaning
CVE-2026-40967, published by Spring on 27 April 2026 and rated high, affects FilterExpressionConverter implementations used through Spring AI's VectorStore. In several converters, keys and values were not properly escaped, which could let them alter the query. It affects applications that pass user-supplied input as a filter expression, in versions 1.0.0 to 1.0.5 and 1.1.0 to 1.1.4. Fixes are in 1.0.6 and 1.1.5.
The advisory is about query injection rather than tenancy as such. It matters here because vector store filters are a common place to enforce tenancy. A filter can look right as a data structure and still change meaning when it is turned into a backend query. So the restriction has to hold in the query that actually runs, not only in the logical expression.
LiteLLM: the cache that compared the wrong key
CVE-2026-89032, published by VulnCheck on 25 September 2026 with a CVSS v4 score of 8.7, describes a tenant isolation bypass in LiteLLM's semantic cache layer before 1.101.0-rc.1. According to the advisory, a mismatch between two helper functions over a metadata key let an authenticated user with a valid virtual key read other tenants' cached responses by sending semantically similar prompts. Cached tool call payloads could also reach a different principal.
A semantic cache is built to reuse answers for prompts that merely resemble each other, so a scope error reaches further than it would in an exact-match cache. A similar prompt is never permission to reuse another tenant's answer. The scope check has to come before similarity matching.
What both ask of you
The mechanisms differ, but the concern is shared: the scope you intend must match the operation your system actually executes. For anyone building on shared infrastructure, that suggests some practical rules:
- Take tenant identity from your authentication layer, never from a prompt, tool argument or metadata alias.
- Keep the tenant restriction outside anything a user can influence, and test the real backend with quotes, escapes, nested expressions and unsupported syntax.
- Use one immutable scope for cache writes, lookups, refresh, invalidation and retries. If a required scope is missing, deny instead of falling back to a shared bucket.
- Check ownership on the returned entry. A key prefix or a hash is not evidence of ownership.
- Restrict candidates to the authorised scope before semantic matching.
- Treat cached tool calls as data. A cache hit does not authorise running a tool.
Where Ten-Grd fits, and where it doesn't
Ten-Grd, our Apache-2.0 Python SDK, is built around this problem. It adds a mandatory tenant restriction to supported selectors, uses the Z3 solver to check whether a selector could include another tenant's data, and validates each result batch before returning it. Anything it can't verify is denied.
It does not patch Spring AI or LiteLLM and has no integration with either, so if you run those products you still need the vendor fixes. Our regression tests draw on these failure families but don't reproduce the affected products. The check applies to a supported selector model under trusted host and adapter assumptions, and your application still needs its own authentication and authorization.