vLLM: Uncontrolled resource consumption via image processing (CVE-2026-116177)
An unauthenticated denial of service vulnerability in vLLM versions 0.10.2 through 0.24.x allows attackers to trigger an out-of-memory condition by submitting decompression-bomb images to the /v1/chat/completions endpoint.
CVE search metadata
CVE search record: CVE-2026-116177. Severity: high. CVSS: 7.5. KEV: no. Product: vLLM (0.10.2 <= version < 0.25.0), vLLM (0.8.5 <= version < 0.31.0). Brief: vLLM: Uncontrolled resource consumption via image processing (CVE-2026-116177). Brief link: https://feed.craftedsignal.io/briefs/2026-10-vllm-resource-exhaustion/
CVE search record: CVE-2026-116180. Severity: high. CVSS: 7.7. KEV: no. Product: vLLM (0.10.2 <= version < 0.25.0), vLLM (0.8.5 <= version < 0.31.0). Brief: vLLM: Uncontrolled resource consumption via image processing (CVE-2026-116177). Brief link: https://feed.craftedsignal.io/briefs/2026-10-vllm-resource-exhaustion/
Triage facts
Revision 1- Applicability
- vLLM (0.10.2 <= version < 0.25.0), vLLM (0.8.5 <= version < 0.31.0)
- Evidence
- Unknown
- Severity
- Low
- Action
- Elevated
- Sources
- 2 public sources
- Freshness
- Revision 1Oct 25, 14:32 UTC
- Vulnerabilities
- 2 CVEs
- Source quality
- A1Likely
What's new
- 1. added coverage for vLLM (0.8.5 <= version < 0.31.0) Oct 25, 14:32 via nvd
vLLM versions 0.10.2 through 0.24.x are vulnerable to uncontrolled resource consumption (CVE-2026-116177). The vulnerability originates in the Nemotron-VL processor configuration, which sets the Pillow library attribute Image.MAX_IMAGE_PIXELS to None. This removes the default safety limit for image processing, allowing an unauthenticated attacker to supply a large, malicious image file to the /v1/chat/completions API endpoint, leading to process memory exhaustion and service denial.
Impact
Successful exploitation results in an OOM (out-of-memory) kill of the vLLM worker process, causing a denial of service for all users of the affected worker instance.
Recommendation
- Upgrade vLLM to version 0.25.0 or later immediately to restore image pixel limits.
- Audit network access to the /v1/chat/completions endpoint to restrict exposure to trusted clients where possible.
- Configure resource limits (cgroups or container-level memory limits) for vLLM workers to mitigate the impact of OOM events.
Immediate actions
Upgrade vLLM to 0.25.0 or later
Upgrade vLLM to version 0.31.0 or later
Threat Hunt
Identify abnormal memory usage spikes or frequent process restarts of vLLM worker containers
Data: Container metrics (memory usage), Orchestrator logs (OOMKill events)
Search application logs for repeated OverflowError exceptions associated with tokenizer decode functions.
Data: Application logs (stderr/stdout from vLLM engine)
Mitigations
Upgrade vLLM to 0.25.0
CVE-2026-116177
Upgrade to vLLM 0.31.0
CVE-2026-116180