CVE-2026-44435: unauthenticated QUIC DoS

CISO Take

This is a denial-of-service bug in quicly, the QUIC (HTTP/3) implementation embedded in the H2O web server: an attacker can send oversized CRYPTO-stream handshake data (>32KB) in a single packet number space and trigger an internal assertion failure that crashes the server process, with no authentication or user interaction required. It is not an AI-specific vulnerability — it lives in the transport layer, not in any model, agent, or inference logic — but any AI product that fronts its inference API, RAG endpoint, or agent gateway with H2O over HTTP/3 inherits this crash-on-demand exposure. The urgency is moderate rather than critical: EPSS sits at 0.00278 (bottom of the exploitation-likelihood distribution), it is not in CISA KEV, and no public exploit or Nuclei template exists yet, so this looks like a quiet infrastructure bug rather than an active campaign. Patch quicly to a version including commit 937d0e9 (or update H2O to a release that bundles it) on any edge server that terminates HTTP/3/QUIC in front of an AI workload, and if HTTP/3 isn't a hard requirement, disabling QUIC on that listener removes the exposure entirely while you schedule the patch.

Sources: NVD GitHub Advisory EPSS ATLAS

What is the risk?

Network-exploitable, unauthenticated, low-complexity (AV:N/AC:L/PR:N/UI:N) and yields a full availability hit (A:H) with no confidentiality or integrity loss — a classic remote crash-DoS profile, correctly scored 7.5/High. Real-world risk is tempered by narrow deployment scope (only servers actually running quicly/H2O with QUIC/HTTP3 enabled are affected), very low EPSS (0.00278), absence from CISA KEV, and no known public PoC or scanner signature. Exploitability is technically trivial once a PoC exists — it requires no AI/ML expertise, just a malformed QUIC handshake — so the low current EPSS mainly reflects lack of published exploit code rather than inherent difficulty.

How does the attack unfold?

Reconnaissance

Adversary scans for internet-facing servers running H2O with QUIC/HTTP3 enabled, potentially fronting an AI inference API or agent gateway.

AML.T0006

Exploitation

Attacker sends a crafted QUIC connection whose CRYPTO-stream handshake data exceeds 32KB in a single packet number space, triggering an internal assertion failure in quicly.

AML.T0049

Impact

The H2O/quicly process crashes, denying availability to whatever AI service (inference API, RAG backend, agent gateway) was served through that edge node.

AML.T0029

Reconnaissance

Adversary scans for internet-facing servers running H2O with QUIC/HTTP3 enabled, potentially fronting an AI inference API or agent gateway.

AML.T0006

Exploitation

Attacker sends a crafted QUIC connection whose CRYPTO-stream handshake data exceeds 32KB in a single packet number space, triggering an internal assertion failure in quicly.

AML.T0049

Impact

The H2O/quicly process crashes, denying availability to whatever AI service (inference API, RAG backend, agent gateway) was served through that edge node.

AML.T0029

How severe is it?

CVSS 3.1

7.5 / 10

EPSS

0.5%

chance of exploitation in 30 days

Higher than 41% of all CVEs

Source: EPSS v3 — FIRST.org

Exploitation Status

No known exploitation

Sophistication

Trivial

What is the attack surface?

AV Network

AC Low

PR None

UI None

S Unchanged

C None

I None

A High

What should I do?

1 step

Upgrade quicly (and any H2O build bundling it) to a version incorporating commit 937d0e9 or later, per the vendor advisory GHSA-2cw9-5673-73gv. If immediate patching isn't feasible, disable HTTP/3/QUIC on the affected listener and fall back to HTTP/2/TLS until the fix is deployed. Add monitoring for repeated H2O/quicly process crashes or assertion-failure log entries as a detection signal, since a crash loop is the primary observable of active exploitation. Inventory any AI-serving infrastructure (inference gateways, RAG proxies, agent API fronts) that terminates HTTP/3 via H2O to confirm exposure before deprioritizing this as generic web-server hygiene.

What does CISA's SSVC say?

Decision Track

Exploitation none

Automatable Yes

Technical Impact partial

Source: CISA Vulnrichment (SSVC v2.0). Decision based on the CISA Coordinator decision tree.

How is it classified?

DoS API Inference AML.T0029 - Denial of AI Service AML.T0049 - Exploit Public-Facing Application

Which compliance frameworks are affected?

This CVE is relevant to:

EU AI Act

Article 15 - Accuracy, robustness and cybersecurity

NIST AI RMF

MEASURE 2.7 - AI system security and resilience is evaluated and documented

OWASP LLM Top 10

LLM10:2025 - Unbounded Consumption

Frequently Asked Questions

What is CVE-2026-44435?

This is a denial-of-service bug in quicly, the QUIC (HTTP/3) implementation embedded in the H2O web server: an attacker can send oversized CRYPTO-stream handshake data (>32KB) in a single packet number space and trigger an internal assertion failure that crashes the server process, with no authentication or user interaction required. It is not an AI-specific vulnerability — it lives in the transport layer, not in any model, agent, or inference logic — but any AI product that fronts its inference API, RAG endpoint, or agent gateway with H2O over HTTP/3 inherits this crash-on-demand exposure. The urgency is moderate rather than critical: EPSS sits at 0.00278 (bottom of the exploitation-likelihood distribution), it is not in CISA KEV, and no public exploit or Nuclei template exists yet, so this looks like a quiet infrastructure bug rather than an active campaign. Patch quicly to a version including commit 937d0e9 (or update H2O to a release that bundles it) on any edge server that terminates HTTP/3/QUIC in front of an AI workload, and if HTTP/3 isn't a hard requirement, disabling QUIC on that listener removes the exposure entirely while you schedule the patch.

Is CVE-2026-44435 actively exploited?

No confirmed active exploitation of CVE-2026-44435 has been reported, but organizations should still patch proactively.

How to fix CVE-2026-44435?

Upgrade quicly (and any H2O build bundling it) to a version incorporating commit 937d0e9 or later, per the vendor advisory GHSA-2cw9-5673-73gv. If immediate patching isn't feasible, disable HTTP/3/QUIC on the affected listener and fall back to HTTP/2/TLS until the fix is deployed. Add monitoring for repeated H2O/quicly process crashes or assertion-failure log entries as a detection signal, since a crash loop is the primary observable of active exploitation. Inventory any AI-serving infrastructure (inference gateways, RAG proxies, agent API fronts) that terminates HTTP/3 via H2O to confirm exposure before deprioritizing this as generic web-server hygiene.

What systems are affected by CVE-2026-44435?

This vulnerability affects the following AI/ML architecture patterns: model serving / inference gateways, agent frameworks (API/edge layer), RAG pipelines (if fronted by H2O/QUIC).

What is the CVSS score for CVE-2026-44435?

CVE-2026-44435 has a CVSS v3.1 base score of 7.5 (HIGH). The EPSS exploitation probability is 0.49%.

What is the AI security impact?

Affected AI Architectures

model serving / inference gatewaysagent frameworks (API/edge layer)RAG pipelines (if fronted by H2O/QUIC)

MITRE ATLAS Techniques

AML.T0029 Denial of AI Service

AML.T0049 Exploit Public-Facing Application

Compliance Controls Affected

EU AI Act: Article 15

NIST AI RMF: MEASURE 2.7

OWASP LLM Top 10: LLM10:2025

What are the technical details?

Original Advisory

Quicly is an IETF QUIC protocol implementation intended primarily for use within the H2O HTTP server. Prior to commit 937d0e9, an assertion failure is raised when the total number of valid handshake messages received over a CRYPTO stream of a single packet number space exceeds 32KB, causing a Denial of Service. This issue has been fixed by commit 937d0e9.

Exploitation Scenario

An adversary scans internet-facing endpoints for HTTP/3-enabled servers and identifies one running H2O with quicly for QUIC termination in front of a company's LLM inference API or agent gateway. They craft a QUIC connection where the total valid handshake data delivered over the CRYPTO stream in a single packet number space exceeds 32KB, which trips an internal assertion in quicly and crashes the server process. Repeating this against the edge node causes a sustained denial of service, taking the AI inference endpoint offline for legitimate users and downstream applications without needing any credentials or prior access.

Weaknesses (CWE)

CWE-400 Uncontrolled Resource Consumption Primary CWE-617 Reachable Assertion Primary CWE-400 Uncontrolled Resource Consumption CWE-617 Reachable Assertion

CWE-400 — Uncontrolled Resource Consumption: The product does not properly control the allocation and maintenance of a limited resource.

[Architecture and Design] Design throttling mechanisms into the system architecture. The best protection is to limit the amount of resources that an unauthorized user can cause to be expended. A strong authentication and access control model will help prevent such attacks from occurring in the first place. The login application should be protected against DoS attacks as much as possible. Limiting the database access, perhaps by caching result sets, can help minimize the resources expended. To further limit the potential for a DoS attack, consider tracking the rate of requests received from users and blocking requests that exceed a defined rate threshold.
[Architecture and Design] Mitigation of resource exhaustion attacks requires that the target system either: The first of these solutions is an issue in itself though, since it may allow attackers to prevent the use of the system by a particular valid user. If the attacker impersonates the valid user, they may be able to prevent the user from accessing the server in question. The second solution is simply difficult to effectively institute -- and even when properly done, it does not provide a full solution. It simply makes the attack require more resources on the part of the attacker. recognizes the attack and denies that user further access for a given amount of time, or uniformly throttles all requests in order to make it more difficult to consume resources more quickly than they can again be freed.

Source: MITRE CWE corpus.