CVE-2026-50559: Quarkus: HTTP auth bypass via %3B/%2F path smuggling

GHSA-qcxp-gm7m-4j5v HIGH PoC AVAILABLE CISA: TRACK*
Published July 29, 2026
CISO Take

Quarkus's path-based HTTP security layer can be bypassed by encoding a semicolon (%3B) to smuggle a matrix parameter past authorization checks, or by encoding a slash/backslash (%2F/%5C) to reach static files that a security policy was supposed to protect. Any application using `quarkus.http.auth.permission` path rules — a very common pattern for gating admin panels, internal APIs, or model/artifact static file serving in AI backends — is exposed with a single crafted curl request and no authentication. There's no CISA KEV listing, no public PoC repo, and no Nuclei template yet, and EPSS sits at a modest 0.46% (top 62nd percentile), so this reads as high-severity-but-not-yet-weaponized rather than an active-exploitation emergency; still, the CVSS 7.5 (confidentiality-only, network, no auth, no user interaction) makes it trivially scriptable once someone builds a scanner signature. Given the package sits behind 2,246 downstream dependents, patch `io.quarkus:quarkus-vertx-http` to 3.20.6.2+ now, and in the interim audit any path-policy-protected endpoint or static resource for reachability via `%3B`, `%2F`, `%5C`, or double-encoded (`%252F`) variants.

Sources: NVD GitHub Advisory EPSS CISA KEV ATLAS

What is the risk?

High severity (CVSS 7.5, confidentiality-only impact) with trivial exploitability — a single unauthenticated HTTP request with URL-encoded characters is sufficient, requiring no AI/ML-specific knowledge. Mitigating factors: not in CISA KEV, no public exploit code or Nuclei template identified, and EPSS is low (0.46%, ~62nd percentile), suggesting opportunistic mass exploitation hasn't started. However, the vulnerability is easy to weaponize once known (it's a two-line curl PoC) and the affected component is a core HTTP security layer used broadly in Java-based backends, including those fronting AI inference services, MLOps consoles, and internal agent tooling. Risk should be treated as high until patched, particularly for any AI-serving app that relies on path-based policies rather than annotation-based (`@RolesAllowed`) security.

How does the attack unfold?

Recon
Attacker fingerprints the target as Quarkus-based and identifies endpoints or static resources gated by quarkus.http.auth.permission path policies.
AML.T0006
Path smuggling
Attacker crafts a request with an encoded semicolon, slash, or backslash so the security layer's partial decoding disagrees with the downstream handler's full decoding.
AML.T0049
Authorization bypass
The mismatched path fails to match the protective policy while still resolving to the protected endpoint or file, so the request proceeds unauthenticated.
Impact
Attacker gains unauthorized access to admin/model-management APIs or protected static resources such as model artifacts and configuration files.

What systems are affected?

Package Ecosystem Vulnerable Range Patched
JAX maven < 3.20.6.2 3.20.6.2
36.3K 2.5K dependents Pushed 2d ago 100% patched ~42d to patch Full package profile →

Do you use JAX? You're affected.

How severe is it?

CVSS 3.1
7.5 / 10
EPSS
0.7%
chance of exploitation in 30 days
Higher than 50% of all CVEs
Exploitation Status
Exploit Available
Exploitation: MEDIUM
Sophistication
Trivial
Exploitation Confidence
medium
CISA SSVC: Public PoC
Public PoC indexed (trickest/cve)
Composite signal derived from CISA KEV, VulnCheck KEV, CISA SSVC, EPSS, Metasploit, Exploit-DB, trickest/cve, Nuclei templates, and inthewild.io exploitation reports.

What is the attack surface?

AV AC PR UI S C I A
AV Network
AC Low
PR None
UI None
S Unchanged
C High
I None
A None

What should I do?

1 step
  1. 1) Patch to io.quarkus:quarkus-vertx-http 3.20.6.2 or later, which replaces pathWithoutMatrixParams() with a fully percent-decoding normalizePath() that also strips null bytes and resolves dot segments. 2) Until patched, inventory all endpoints and static resources protected by quarkus.http.auth.permission path rules and verify they cannot be reached via %3B, lowercase %3b, %2F, %5C, or double-encoded %252F variants. 3) Where feasible, move critical endpoints to annotation-based security (@RolesAllowed/@Authenticated) as a stopgap, since JAX-RS routing is not affected by the slash/backslash vectors. 4) Add WAF/reverse-proxy rules that reject or fully normalize encoded semicolons, slashes, and backslashes in request paths before they reach the application. 5) Detection: monitor access logs for %3B, %3b, %2F, %5C, or %25 sequences in paths matching protected route prefixes, especially followed by 2xx responses.

What does CISA's SSVC say?

Decision Track*
Exploitation poc
Automatable Yes
Technical Impact partial

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

How is it classified?

Which compliance frameworks are affected?

This CVE is relevant to:

EU AI Act
Article 15 - Accuracy, robustness and cybersecurity
ISO 42001
A.6.2.5 - Third-party and customer requirements to AI system security
NIST AI RMF
MEASURE 2.7 - AI system security and resilience are evaluated and documented

Frequently Asked Questions

What is CVE-2026-50559?

Quarkus's path-based HTTP security layer can be bypassed by encoding a semicolon (%3B) to smuggle a matrix parameter past authorization checks, or by encoding a slash/backslash (%2F/%5C) to reach static files that a security policy was supposed to protect. Any application using `quarkus.http.auth.permission` path rules — a very common pattern for gating admin panels, internal APIs, or model/artifact static file serving in AI backends — is exposed with a single crafted curl request and no authentication. There's no CISA KEV listing, no public PoC repo, and no Nuclei template yet, and EPSS sits at a modest 0.46% (top 62nd percentile), so this reads as high-severity-but-not-yet-weaponized rather than an active-exploitation emergency; still, the CVSS 7.5 (confidentiality-only, network, no auth, no user interaction) makes it trivially scriptable once someone builds a scanner signature. Given the package sits behind 2,246 downstream dependents, patch `io.quarkus:quarkus-vertx-http` to 3.20.6.2+ now, and in the interim audit any path-policy-protected endpoint or static resource for reachability via `%3B`, `%2F`, `%5C`, or double-encoded (`%252F`) variants.

Is CVE-2026-50559 actively exploited?

Proof-of-concept exploit code is publicly available for CVE-2026-50559, increasing the risk of exploitation.

How to fix CVE-2026-50559?

1) Patch to `io.quarkus:quarkus-vertx-http` 3.20.6.2 or later, which replaces `pathWithoutMatrixParams()` with a fully percent-decoding `normalizePath()` that also strips null bytes and resolves dot segments. 2) Until patched, inventory all endpoints and static resources protected by `quarkus.http.auth.permission` path rules and verify they cannot be reached via `%3B`, lowercase `%3b`, `%2F`, `%5C`, or double-encoded `%252F` variants. 3) Where feasible, move critical endpoints to annotation-based security (`@RolesAllowed`/`@Authenticated`) as a stopgap, since JAX-RS routing is not affected by the slash/backslash vectors. 4) Add WAF/reverse-proxy rules that reject or fully normalize encoded semicolons, slashes, and backslashes in request paths before they reach the application. 5) Detection: monitor access logs for `%3B`, `%3b`, `%2F`, `%5C`, or `%25` sequences in paths matching protected route prefixes, especially followed by 2xx responses.

What systems are affected by CVE-2026-50559?

This vulnerability affects the following AI/ML architecture patterns: model serving, agent frameworks, MLOps platforms.

What is the CVSS score for CVE-2026-50559?

CVE-2026-50559 has a CVSS v3.1 base score of 7.5 (HIGH). The EPSS exploitation probability is 0.67%.

What is the AI security impact?

Affected AI Architectures

model servingagent frameworksMLOps platforms

MITRE ATLAS Techniques

AML.T0006 Active Scanning
AML.T0049 Exploit Public-Facing Application

Compliance Controls Affected

EU AI Act: Article 15
ISO 42001: A.6.2.5
NIST AI RMF: MEASURE 2.7

What are the technical details?

Original Advisory

Quarkus HTTP path-based authorization policies can be bypassed using encoded semicolons (%3B) to smuggle matrix parameters past the security layer, and using encoded slashes (%2F) or backslashes (%5C) to access protected static resources. This is a distinct issue from CVE-2026-39852, which addressed only literal semicolon stripping. ### Technical Details The security layer (AbstractPathMatchingHttpSecurityPolicy) normalizes request paths using Vert.x's normalizedPath(), which only decodes unreserved RFC 3986 characters (letters, digits, -, ., _, ~). It then strips matrix parameters by looking for literal ; characters. This creates two mismatches: 1. Encoded semicolons (`%3B`): Since `%3B` is not decoded by normalizedPath(), the matrix parameter stripping in pathWithoutMatrixParams() never sees it. The encoded semicolon and everything after it become part of the path segment, causing policy matching to fail. This affects all path-policy-protected endpoints. 2. Static resource path mismatch: Static resource handlers (StaticHandlerImpl, FileSystemStaticHandler) perform full percent-decoding via URIDecoder.decodeURIComponent() and backslash-to-slash conversion before filesystem resolution. Reserved characters like `%2F` (slash) and `%5C` (backslash) that survive the security layer's partial decoding are fully decoded before file serving. REST endpoints using Quarkus REST (RESTEasy Reactive) are not affected by the `%2F/%5C` vectors because the routing layer also uses `normalizedPath()` — both security and routing agree on the path, so no mismatch exists. ### Attack Vectors Encoded semicolon (matrix parameter smuggling), affects all path-policy-protected endpoints: - `/api/admin%3Bbypass=true/data`: security sees this as a single segment `admin%3Bbypass=true`, which does not match the `/api/admin/* policy`. The request passes through unauthenticated. - `/api/secret%3b/data`: same mechanism with lowercase hex digit. Encoded slash/backslash on static resources, affects static files behind path policies: - `/static-secret%2Fhtml`, security does not match /static-secret.html policy; static handler decodes `%2F` to `/` and may resolve the file. - `/static-secret%5Chtml `. static handler decodes `%5C` to `\`, then converts to `/`. Double encoding, affects static resources: - `/secret%252Fconfidential.html`, first decode by normalizedPath() turns `%25` into `%`, producing `%2F`. Static handler's second decode turns `%2F` into `/`. The following vectors were investigated and confirmed not exploitable: - Unreserved character encoding (`/api/adm%69n/data`): `normalizedPath()` decodes these. Both security and routing see `/api/admin/data`. - Null byte injection (`/api/admin%00/data`): `%00` is not decoded by `normalizedPath()`. - Encoded dot segments (`/api/%2e%2e/secret/data`): Period is unreserved, so `%2e` is decoded to `.` by `normalizedPath()`, then `removeDots()` normalizes `..` segments. - REST endpoint bypass via `%2F/%5C`: Routing uses the same `normalizedPath()` as security. The encoded slash/backslash doesn't match any route. ### Root Cause `pathWithoutMatrixParams()` operates on the partially-decoded output of `normalizedPath()`, where reserved characters remain encoded. It searches for literal ; but never sees `%3B.` The fix (`normalizePath()`) performs full percent-decoding in a loop before stripping matrix parameters, removing null bytes, normalizing backslashes, and resolving dot segments, aligning the security layer's view of the path with what downstream handlers resolve. ### Impact - Unauthenticated access to endpoints protected by `quarkus.http.auth.permission` path-based policies via `%3B` smuggling - Static resource exposure by bypassing path policies on protected files via `%2F/%5C` - Applications using annotation-based security (`@RolesAllowed`, `@Authenticated`) on JAX-RS resources without path-based policies are not affected by the `%2F/%5C` vectors, but may still be affected by `%3B` if path policies coexist ### Proof of Concept ``` # Encoded semicolon bypass — works on any path-policy-protected endpoint # Security sees "/api/admin%3Bbypass=true/data", doesn't match /api/admin/* policy curl -v http://target/api/admin%3Bbypass=true/data # Encoded semicolon on authenticated endpoint curl -v http://target/api/secret%3b/data # Static resource bypass via encoded slash (if static file behind path policy) curl -v http://target/static-secret%2Fhtml # Static resource bypass via encoded backslash curl -v http://target/static-secret%5Chtml ```

Exploitation Scenario

An attacker fingerprints a Quarkus-based AI platform (e.g., via response headers or error pages) and identifies that `/api/admin/*` — the model-management and configuration API — is gated by a path-based `quarkus.http.auth.permission` policy rather than per-endpoint annotations. The attacker requests `/api/admin%3Bx=1/models` instead of `/api/admin/models`; the security layer treats `admin%3Bx=1` as an unrecognized segment that doesn't match the `/api/admin/*` policy and lets the request through unauthenticated, while the downstream handler still resolves it to the real admin models endpoint. Separately, the attacker finds that internal model weights or a compliance config file is served statically behind `/internal-models.html`; requesting `/internal-models%2Fhtml` bypasses the path policy while the static file handler fully decodes `%2F` back to `/` and serves the protected file, resulting in unauthorized access to proprietary model artifacts or configuration data.

Weaknesses (CWE)

CWE-178 — Improper Handling of Case Sensitivity: The product does not properly account for differences in case sensitivity when accessing or determining the properties of a resource, leading to inconsistent results.

  • [Architecture and Design] Avoid making decisions based on names of resources (e.g. files) if those resources can have alternate names.
  • [Implementation] Assume all input is malicious. Use an "accept known good" input validation strategy, i.e., use a list of acceptable inputs that strictly conform to specifications. Reject any input that does not strictly conform to specifications, or transform it into something that does. When performing input validation, consider all potentially relevant properties, including length, type of input, the full range of acceptable values, missing or extra inputs, syntax, consistency across related fields, and conformance to business rules. As an example of business rule logic, "boat" may be syntactically valid because it only contains alphanumeric characters, but it is not valid if the input is only expected to contain colors such as "red" or "blue." Do not rely exclusively on looking for malicious or malformed inputs. This is likely to miss at least one undesirable input, especially if the code's environment changes. This can give attackers enough room to bypass the intended validation. However, denylis

Source: MITRE CWE corpus.

CVSS Vector

CVSS:3.1/AV:N/AC:L/PR:N/UI:N/S:U/C:H/I:N/A:N

Timeline

Published
July 29, 2026
Last Modified
August 13, 2026
First Seen
July 29, 2026

Related Vulnerabilities