vulnerability / NVI-00012185 / Revision 2
vLLM: Mirrored multimodal IPC caches desync after a rejected request — a later request reusing the same media hash trips a receiver assertion in the engine core
GitHub lists this reviewed advisory for vllm. Published version ranges and source prerequisites are shown below; this is not an inspection of a visitor’s application.
This report has not inspected your inventory, configuration, deployed build, reachability or evidence. A catalog listing and a severity score do not establish exposure or compromise.
01 / Meaning
Why it matters
Published affected context helps an owner decide what evidence to gather. This report has not examined your application.
6.5 medium
GitHub source severity; exploitability and local consequence require separate evidence.
As of · GitHub Advisory Database
0.427%
FIRST EPSS model v2026.06.15; exploitation probability over 30 days, percentile 0.37025. Missing scores are unknown.
As of · FIRST EPSS
02 / Applicability
The affected context
Affected versions, configuration prerequisites and known exclusions are public. Confirm each one against the deployed service.
vllm
source listed< 0.28.0
## Affected - **Ecosystem / package:** pip / `vllm` - **Affected versions:** vLLM ≤ 0.25.1 (confirmed on 0.25.1, commit `752a3a504485`). The lower bound predates 0.25.1; maintainers can confirm how far back the mirrored sender/receiver cache protocol reaches. ## Summary vLLM's default multimodal cache (`mm_processor_cache_type="lru"`) mirrors state across two processes: the frontend (P0) holds only metadata (`MultiModalProcessorSenderCache`) while the engine core (P1) holds the real payload (`MultiModalReceiverCache`). The design invariant is that `get_and_update()` runs on P0 and P1 in lockstep for every request, so eviction order stays mirrored and P0 can answer "is this cached in P1?" without talking to P1. That invariant breaks when a request is **rejected after P0 has rendered and hashed the multimodal input** (populating the P0 cache) **but before P1 receives the item** — for example, an oversized chat prompt rejected on `max_model_len` *after* rendering. P0 now believes the media is cached while P1 never got it. A later request reusing the same media hash gets a P0 hit, so P0 sends `Non
03 / Response
What to do next
- investigate
Identify the exact component and deployed version, examine the linked publisher’s conditions, and retain evidence of any mitigation.
Confirm current vendor support, compatibility and deployment evidence. This report has not tested a fix in your system. FIRST EPSS
A fixed release is not deployment evidence.
Check current vendor support and compatibility. After making a change, verify the deployed build and relevant runtime conditions. Keep the evidence, owner and time. An updated lockfile alone does not prove what is running.
04 / Limits
What this report cannot establish
- Customer inventory, deployment, reachability and exploitation have not been inspected.
05 / Provenance
Sources you can inspect
Each assertion retains its source. Mirrored records may repeat one underlying publisher; additional links are not independent corroboration.
- GHSA-ph3r-5jfg-f84f ↗
GitHub Advisory Database · vendor advisory
Observed . GitHub Advisory Database; CC-BY-4.0. Each advisory links its original attribution. Null normalizes fields.
- CVE-2026-105753 ↗
FIRST EPSS · official catalog
Observed . FIRST EPSS; probability of exploitation in the next 30 days, not local compromise. Dataset date/model retained; FIRST services terms apply.
Public revision history
Only published revisions appear here. Routine fetches do not create a material-change event.
- Revision 2 ·
Dated EPSS estimate updated; this is not a new disclosure or proof of exploitation.
- Revision 1 ·
First indexed by Null; source disclosure date is shown separately.