Vigilbase CVEConnect your agent

vLLM leaks a heap address when PIL throws an error

Action requiredNot in CISA KEVMetasploit moduleCVE-2026-22778CWE-532

Not confirmed exploited. 1 of 8 rule sources name it. Our CRS test: partly blocked.

Edge rules
1/8
Our test
Partly blocked
CVSS
9.8
EPSS
18.3% Top 3%
Posts, 14d
02026-10-22: 0 posts2026-10-23: 0 posts2026-10-24: 0 posts2026-10-25: 0 posts2026-10-26: 0 posts2026-10-27: 0 posts2026-10-28: 0 posts2026-10-29: 0 posts2026-10-30: 0 posts2026-10-31: 0 posts2026-11-01: 0 posts2026-11-02: 0 posts2026-11-03: 0 posts2026-11-04: 0 posts
Published
2026-02-02
vLLM is an inference and serving engine for large language models (LLMs). From 0.8.3 to before 0.14.1, when an invalid image is sent to vLLM's multimodal endpoint, PIL throws an error. vLLM returns this error to the client, leaking a heap… More

vLLM is an inference and serving engine for large language models (LLMs). From 0.8.3 to before 0.14.1, when an invalid image is sent to vLLM's multimodal endpoint, PIL throws an error. vLLM returns this error to the client, leaking a heap address. With this leak, we reduce ASLR from 4 billion guesses to ~8 guesses. This vulnerability can be chained a heap overflow with JPEG2000 decoder in OpenCV/FFmpeg to achieve remote code execution. This vulnerability is fixed in 0.14.1.

Action requiredLast checked 2026-11-03

The risk

It is reachable over the network without credentials or any user interaction (CVSS 9.8). A Metasploit module exists, so exploitation needs no custom work. EPSS ranks it in the top 3% of CVEs most likely to be exploited in the next 30 days. Exploitation in the wild is not established by CISA KEV.

Who is affected

Teams running vllm-project vllm (>= 0.8.3, < 0.14.1). Prioritise internet-facing instances.

Next steps

  1. Apply the vendor's fix; the advisory lists the affected and fixed versions.
  2. Keep a WAF rule in place while you patch: F5 ships a rule you may need to switch on. A WAF rule reduces exposure; it does not replace the patch. Edge coverage
  3. Check which WAF sits in front of your site and whether it covers this CVE. Check your domain

Attack traffic on Cloudflare

Cloudflare Radar
  • Remote Code Execution1.8% of mitigated layer 7 attacks, 7 days2026-10-06: 5.7%2026-10-07: 3.8%2026-10-08: 10.6%2026-10-09: 6.9%2026-10-10: 11.8%2026-10-11: 4.7%2026-10-12: 2.4%2026-10-13: 1.9%2026-10-14: 1.3%2026-10-15: 1.5%2026-10-16: 1.6%2026-10-17: 1.3%2026-10-18: 2.2%2026-10-19: 1.3%2026-10-20: 2.6%2026-10-21: 2.3%2026-10-22: 5.5%2026-10-23: 1.6%2026-10-24: 1.1%2026-10-25: 3.2%2026-10-26: 3.2%2026-10-27: 3%2026-10-28: 2.8%2026-10-29: 2.2%2026-10-30: 1.8%2026-10-31: 1.6%2026-11-01: 1.3%2026-11-02: 1.3%2026-11-03: 1.4%
  • Command Injection5.5% of mitigated layer 7 attacks, 7 days2026-10-06: 10.1%2026-10-07: 10.1%2026-10-08: 15.8%2026-10-09: 14.2%2026-10-10: 28.8%2026-10-11: 12.2%2026-10-12: 6.8%2026-10-13: 5.7%2026-10-14: 5.8%2026-10-15: 5.4%2026-10-16: 5.1%2026-10-17: 5.7%2026-10-18: 5.5%2026-10-19: 5.1%2026-10-20: 11.3%2026-10-21: 8.5%2026-10-22: 12.8%2026-10-23: 4.4%2026-10-24: 5.2%2026-10-25: 9.2%2026-10-26: 8.5%2026-10-27: 7.2%2026-10-28: 7.6%2026-10-29: 4.6%2026-10-30: 7.5%2026-10-31: 5.7%2026-11-01: 4.6%2026-11-02: 4.4%2026-11-03: 4.6%
  • Rules for named CVEs5.9% of mitigated layer 7 attacks, 7 days2026-10-06: 0.6%2026-10-07: 1.8%2026-10-08: 3%2026-10-09: 6.3%2026-10-10: 1.9%2026-10-11: 0.8%2026-10-12: 0.6%2026-10-13: 0.6%2026-10-14: 4.2%2026-10-15: 2.3%2026-10-16: 1.3%2026-10-17: 1.6%2026-10-18: 1.1%2026-10-19: 3.2%2026-10-20: 4.7%2026-10-21: 5.7%2026-10-22: 5.5%2026-10-23: 9.8%2026-10-24: 7%2026-10-25: 5%2026-10-26: 7.3%2026-10-27: 8.1%2026-10-28: 10.1%2026-10-29: 0.9%2026-10-30: 4.1%2026-10-31: 7.5%2026-11-01: 8.7%2026-11-02: 2.4%2026-11-03: 9.1%

Radar groups Cloudflare's WAF managed rules by category; it does not publish counts per CVE. These shares show how much attack traffic targets this kind of flaw across Cloudflare's network (28-day trend). They are not compromised systems, and seeing few attempts on Cloudflare does not prove an absence elsewhere. Radar data as of 2026-11-03. Radar ↗

Check your exposure with an agent

Read-only

Copy this into an agent with access to your infrastructure. It asks for an evidence-based, read-only assessment: affected systems, exposure, compensating controls and what is still unknown. It does not scan or change anything by itself.

Prompt12 lines
Assess whether the infrastructure available to you is affected by CVE-2026-22778 (vLLM leaks a heap address when PIL throws an error). Perform a read-only assessment within your authorized access. Do not send exploit traffic, scan third-party systems, change configuration or restart services.

1. Verify the CVE record at https://www.cve.org/CVERecord?id=CVE-2026-22778 and the vendor advisory. Establish affected and fixed versions, including vendor backports. The record lists: vllm-project vllm: affected >= 0.8.3, < 0.14.1. Cite sources and retrieval dates; flag anything you cannot verify.
2. Inventory deployments of vllm-project vllm: asset, environment, owner where known, installed version or build, and how you confirmed it (package metadata, a version endpoint, image tags). An image tag alone does not prove the running version.
3. Identify which deployments are reachable from the internet or untrusted networks, and through which paths (direct origin, load balancer, CDN or WAF).
4. Assess compensating controls separately from patch status. Current edge WAF coverage for this CVE: F5: rule exists, check whether it is enabled. Live data: https://cve.vigilbase.com/api/v1/cves/CVE-2026-22778. Check whether the rule is enabled and in block mode on each path, and whether any path bypasses the WAF. A WAF rule is not a patch.
5. Check logs on exposed instances for requests matching the public exploit, if your access allows.

Return a short summary and a table: asset | version/build | assessment (affected, fixed, not applicable, unknown) | exposure | compensating control | evidence | next action. List what you could not assess and the access needed to resolve it. Do not infer absence of compromise from patch status or a lack of alerts.

Remediation guidance: Apply the vendor's fix; the advisory lists the affected and fixed versions. Keep a WAF rule in place while you patch: F5 ships a rule you may need to switch on. A WAF rule reduces exposure; it does not replace the patch.
More: https://cve.vigilbase.com/cve/CVE-2026-22778

Sources

Cited
CVE record

CVSS 3.1 9.8 (critical). Weakness: CWE-532.

CVE-2026-22778 ↗NVD ↗

Exploit evidence

Metasploit has a module. A nuclei template exists, which our lab and benchmark replay.

Metasploit: vLLM Multimodal Heap-Address Information Leak Scanner ↗

CISA KEV

Not in the KEV catalog (catalog 2026.11.03, checked 2026-11-04). Absence from KEV is not evidence that it cannot be exploited.

CISA KEV catalog ↗

EPSS

18.3% probability of exploitation activity in the next 30 days, higher than 97% of CVEs (2026-11-02).

FIRST EPSS ↗

WAF vendor statements

F5 names it in a changelog. Quotes and dates below.

Vendor statements

Ask Vigil

AI · answers from this page
Ask about CVE-2026-22778: who's covered, what to patch first, how vendors compare. I cite sources and won't help with attacks.

Severity, decoded

CVSS 3.1
Critical · 9.8 / 10

Exploitable over the network, with no privileges and no user interaction, leading to full loss of confidentiality, integrity and availability.

Over the internetNo login neededNo user actionWorks reliablyFull compromise
How the score is built8 metrics
  • Attack vectorAn attacker can reach it over the internet; no access to your network needed.
    Network
  • Attack complexityNo special conditions are needed; the attack works reliably.
    Low
  • Privileges requiredNo account or login is needed.
    None
  • User interactionNo one has to click or open anything; the attack works on its own.
    None
  • ScopeThe damage stays within the vulnerable component.
    Unchanged
  • Confidentiality impactAn attacker can read all data the affected component holds, or the most sensitive parts of it.
    High
  • Integrity impactAn attacker can change any data or files the affected component protects.
    High
  • Availability impactAn attacker can take the component down completely or keep it from serving anyone.
    High
CVSS:3.1/AV:N/AC:L/PR:N/UI:N/S:U/C:H/I:H/A:H
CISA's assessmentSSVC, from the CVE record
ExploitationNone

CISA has seen no public exploit and no attacks.

AutomatableYes

Attackers can script the attack end to end and hit many targets at once.

Technical impactTotal

A successful attack gives full control over the affected software or all of its data.

Public exploits

2 found
1 Metasploit0 Exploit-DB1 Scanner template0 GitHub PoCs

Links are to third-party code: never run it outside an isolated lab.

Timeline

  1. 2026-02-02
    CVE published
    CVE List
  2. 2026-02-02
    Metasploit module published
    vLLM Multimodal Heap-Address Information Leak Scanner
  3. 2026-09-16
    F5 ships a rule
    vLLM Information Disclosure
  4. 2026-11-02
    Our test: OWASP Core Rule Set blocked part of it
    paranoia level 1, anomaly threshold 5, default CRS settings
  5. 2026-11-02
    Our test: OWASP Core Rule Set blocked part of it
    paranoia level 2, anomaly threshold 5, default CRS settings

Social

2026-10-22: 0 posts2026-10-23: 0 posts2026-10-24: 0 posts2026-10-25: 0 posts2026-10-26: 0 posts2026-10-27: 0 posts2026-10-28: 0 posts2026-10-29: 0 posts2026-10-30: 0 posts2026-10-31: 0 posts2026-11-01: 0 posts2026-11-02: 0 posts2026-11-03: 0 posts2026-11-04: 0 posts

No posts seen yet on X, Bluesky, Mastodon, Hacker News or GitHub.

Vendor statements

Quoted
CloudflareWAF Managed RulesNo public claim

We found no public statement from this vendor naming this CVE. That is not proof of no protection: generic rules may still catch it. Each row gives the rule ID and its default action. The feed starts in January 2025, so older rules are missing.

AWSAWS WAF Managed RulesNo public claim

We found no public statement from this vendor naming this CVE. That is not proof of no protection: generic rules may still catch it. AWS often describes a vulnerability without naming its CVE (for example "Log4j"). We count only entries that name the CVE, so AWS coverage is understated.

Microsoft AzureFront Door and Application Gateway WAF (DRS)No public claim

We found no public statement from this vendor naming this CVE. That is not proof of no protection: generic rules may still catch it. The rule tables name CVEs but carry no dates, so Azure is left out of every time-to-rule figure.

Google CloudCloud Armor preconfigured WAF rulesNo public claim

We found no public statement from this vendor naming this CVE. That is not proof of no protection: generic rules may still catch it. Google names few CVEs: the release notes start in May 2024 and the rules reference lists CVEs without dates. Its cve-canary rules are opt-in.

FastlyNext-Gen WAF virtual patchesBefore feed history

The CVE predates the oldest entry in this vendor's public feed, so a rule may exist that we cannot see.

F5Attack signatures (Distributed Cloud WAF)1 quoteRule published

First rule 2026-09-16, 226 days after the CVE was published.

2026-09-16New rule200009381Source
Added: signature 200009381 "vLLM Information Disclosure" (Information Leakage, risk High), systems: Other Web Server. References: CVE-2026-22778.
AkamaiApp & API Protector (Adaptive Security Engine)No public source

This vendor publishes no machine-readable rule changelog we can check.

OWASP CRSCore Rule Set (open source, used inside many WAFs)No public claim

We found no public statement from this vendor naming this CVE. That is not proof of no protection: generic rules may still catch it. CRS rarely names CVEs: it relies on generic detections. Our own lab measures what it actually blocks.