Loading vulnerability…
CVE-2026-49468 is a critical authentication bypass vulnerability in LiteLLM, an AI gateway and proxy for LLM APIs, affecting versions prior to 1.84.0. The flaw is caused by improper handling of the HTTP Host header during route determination in the proxy authentication layer. Specifically, the auth logic derived the effective route from request.url.path in get_request_route() within litellm/proxy/auth/auth_utils.py, while Starlette reconstructs that path using the supplied Host header. By crafting the Host header, an attacker can cause the authentication layer to evaluate a different route from the one ultimately dispatched by FastAPI. Under specific conditions, this route mismatch can allow unauthenticated access to protected management routes.
Mallory correlates every CVE against your assets, your vendors, and active adversary campaigns. Know which vulnerabilities matter for you, not just which ones are loud.
What it means. What to do now. Patch path, mitigations, and the assume-compromise checklist.
What an attacker gets, and what they’ve been doing with it.
If you can’t patch tonight, do this now.
Patch, then assume compromise.
1 valid exploit after Mallory filtered fakes, detection scripts, and README-only repos.
This repository is a compact, self-contained PoC for CVE-2026-49468 affecting the BerriAI LiteLLM proxy before 1.84.0. The main exploit is exploit.py, a stdlib-only Python script that uses http.client with skip_host=True to send a verbatim malformed Host header ('evil/?') to the real target host. The exploit abuses route confusion in LiteLLM auth logic: authentication/authorization decisions are derived from request.url.path, which Starlette reconstructs from the client-controlled Host header, while FastAPI routing still uses the real ASGI path. By injecting '/?' into Host, the auth layer sees '/' (a public route), causing both authentication and route-based authorization checks to short-circuit. Capabilities implemented in exploit.py include: vulnerability checking via baseline vs bypassed requests to /user/list; unauthenticated minting of a valid API key through POST /key/generate; unauthenticated user creation via POST /user/new; unauthenticated inference via POST /chat/completions; enumeration/disclosure via GET /v1/models, /model/info, /spend/logs, /settings, and /get/config/callbacks; and a generic raw request mode for arbitrary methods/paths. The exploit does not provide code execution payloads and the included analysis explicitly states that an RCE chain through MCP server creation is not reachable on the tested vulnerable version because inline PROXY_ADMIN checks still block those paths. Repository structure: README.md documents the vulnerability, impact, reproduction, and detection guidance; ANALYSIS.md contains deeper root-cause and reachability analysis; exploit.py is the operational PoC; lab/config.yaml and lab/docker-compose.yml provide a reproducible environment with vulnerable LiteLLM v1.83.14-stable on port 4000, patched v1.84.0 on port 4001, and a PostgreSQL backend. Overall, this is a real exploit repository with an operational pre-auth web/network auth-bypass PoC and a lab for validation, not merely a detector or write-up.
Products and vendors Mallory has correlated with this vulnerability. Open in Mallory to drill down to specific CPE configurations and version ranges.
Vendor-confirmed product mapping. Mallory continuously reconciles this list against your asset inventory.
9 sources tracked across advisories, community write-ups, and news. New activity surfaces here as Mallory finds it.
An authentication bypass vulnerability in LiteLLM caused by Host Header Injection, affecting versions prior to 1.84.0.
A critical authentication bypass vulnerability in LiteLLM caused by Host header injection, allowing unauthenticated attackers to reach protected management routes under affected deployment conditions.
A critical authentication bypass vulnerability in LiteLLM caused by improper Host header handling, allowing route confusion between authentication checks and FastAPI processing and enabling unauthorized access to sensitive management endpoints.
Query your assets running an affected version, and investigate the blast radius.
Every observed campaign linking this CVE to a named adversary.
Malware families riding this exploit, with evidence and IOCs.
YARA, Sigma, Snort, and vendor rules, auto-deployed to your SIEM.
Cross-references every affected SKU, including bundled OEM variants.
Community discussion across Reddit, Mastodon, and other social sources.