Skip to main content
Vulnerability Database/CVE-2026-65104

CVE-2026-65104: NVIDIA Megatron-LM Information Disclosure

CVE-2026-65104 is an information disclosure vulnerability in NVIDIA Megatron-LM affecting inference operations through hash collision attacks. This article covers technical details, impact analysis, and mitigation.

Published:

Disclaimer: This content was generated using AI. While we strive for accuracy, please verify critical information with official sources.

Sources: National Vulnerability Database, CISA KEV, Exploit-DB

CVE-2026-65104 Overview

NVIDIA Megatron-LM contains a vulnerability in its inference component where an attacker can force hash collisions in the cache used for inference requests. The flaw affects all supported platforms and stems from the use of a cryptographically weak hashing algorithm [CWE-328] for cache keying. Successful exploitation can lead to data tampering and partial information disclosure between inference requests. The issue impacts the integrity of model outputs served from the shared cache and may expose fragments of other users' request data in multi-tenant inference deployments.

Critical Impact

An authenticated network attacker can craft inputs that collide with cached inference entries, allowing tampering of returned results and partial disclosure of cached request data.

Affected Products

  • NVIDIA Megatron-LM (all platforms)
  • Inference cache subsystem
  • Multi-tenant deployments serving Megatron-LM inference

Discovery Timeline

  • 2026-10-13 - CVE-2026-65104 published to the National Vulnerability Database (NVD)
  • 2026-10-13 - Last updated in NVD database

Technical Details for CVE-2026-65104

Vulnerability Analysis

The vulnerability resides in the inference request caching layer of NVIDIA Megatron-LM. The cache uses a weak hashing algorithm to derive keys from request content, enabling an attacker who can submit inference requests to deliberately produce collisions. When a collision occurs, a crafted request can map to a cache entry populated by a different request, causing the cached response of one request to be returned for another.

The attack produces two outcomes. First, integrity is lost because the attacker can tamper with the data returned to subsequent requests served from the poisoned cache entry. Second, confidentiality is partially broken because cached response fragments associated with other requests may be returned to the attacker. The high attack complexity reflects the computational effort to produce usable collisions, while the scope change indicates impact beyond the attacker's own session.

Root Cause

The root cause is the selection of a weak or broken cryptographic hashing algorithm for cache key generation [CWE-328]. Algorithms vulnerable to collision attacks cannot provide the collision resistance required when the hash input is attacker-controlled. Attackers who understand the hashing scheme can precompute inputs that map to the same cache slot as targeted entries.

Attack Vector

Exploitation requires network access and low privileges, consistent with an authenticated user of an inference endpoint. The attacker submits inference requests crafted to collide with hashes of other cached entries. No user interaction is required. Because the cache is shared across inference requests, the impact crosses the attacker's security scope and affects other consumers of the inference service.

The vulnerability is described in prose only; no verified proof-of-concept code is available. Technical details are published in the NVIDIA Product Security advisory and the NVD entry for CVE-2026-65104.

Detection Methods for CVE-2026-65104

Indicators of Compromise

  • Inference requests from a single client producing unusually high rates of distinct payloads that share identical cache key hashes.
  • Returned inference responses that do not correspond to the submitted prompt content or model parameters.
  • Cache metrics showing anomalous hit rates for newly submitted, previously unseen request payloads.

Detection Strategies

  • Instrument the Megatron-LM inference cache to log cache key, input digest, and request origin, then alert on collisions between unrelated tenants.
  • Baseline cache lookup patterns per client and flag deviations consistent with collision probing.
  • Correlate inference server logs with upstream API gateway telemetry to attribute suspicious request patterns to specific identities.

Monitoring Recommendations

  • Monitor inference endpoints for repeated requests exhibiting near-identical hashed keys but divergent content.
  • Track response consistency by sampling identical requests and alerting when outputs diverge unexpectedly across time.
  • Forward inference service logs to a centralized analytics platform to retain evidence for investigation.

How to Mitigate CVE-2026-65104

Immediate Actions Required

  • Review the NVIDIA Product Security advisory 5881 and apply the fixed Megatron-LM release when published.
  • Restrict inference endpoints to trusted, authenticated clients and enforce per-tenant rate limiting on request submission.
  • Audit multi-tenant Megatron-LM deployments for shared cache usage and isolate caches per tenant where feasible.

Patch Information

NVIDIA publishes remediation guidance for CVE-2026-65104 in the NVIDIA product-security repository entry for advisory 5881. Upgrade to the Megatron-LM version identified in that advisory once available. Additional record details are tracked in the CVE.org record for CVE-2026-65104.

Workarounds

  • Disable the inference request cache if the deployment can tolerate the latency impact until a patched release is applied.
  • Partition the inference cache per tenant or per authenticated principal to prevent cross-tenant collision effects.
  • Add a server-side secret salt to cache key derivation so attackers cannot precompute colliding inputs.
bash
# Configuration example: disable inference cache pending patch
export MEGATRON_INFERENCE_CACHE_ENABLED=false

# Or scope cache keys per authenticated tenant
export MEGATRON_INFERENCE_CACHE_KEY_NAMESPACE="tenant:${TENANT_ID}"

Disclaimer: This content was generated using AI. While we strive for accuracy, please verify critical information with official sources.

Sources: National Vulnerability Database, CISA KEV, Exploit-DB

Experience the Most Advanced Cybersecurity Platform

See how the world’s most intelligent, autonomous cybersecurity platform can protect your organization today and into the future.