CVE-2026-59822

HIGH(8.2)KEV

BerriAI LiteLLM Improper Authentication Vulnerability

Description

CVE-2026-59822 is a high-severity improper authentication vulnerability in LiteLLM, the open-source proxy server and AI gateway used to call LLM APIs in OpenAI-compatible format. Prior to version 1.84.0, LiteLLM's MCP Streamable HTTP endpoint let an unauthenticated attacker supply a fabricated Authorization header that triggered an OAuth2 passthrough fallback path; instead of rejecting the request, the code replaced the failed LiteLLM key validation with an empty UserAPIKeyAuth() object, so requests reached MCP tooling without any valid LiteLLM key. Because MCP tooling exposes the models, connectors and internal capabilities an AI gateway fronts, this authentication bypass vulnerability gives unauthorised access to an organisation's LLM infrastructure. CVE-2026-59822 carries a CVSS score of 8.2, is fixed in LiteLLM 1.84.0, and is listed in the CISA Known Exploited Vulnerabilities (KEV) catalog.

KEV Information

Vendor
BerriAI
Product
LiteLLM
Date Added
September 2, 2026
Due Date
September 16, 2026
Required Action
Apply mitigations in accordance with vendor instructions, ensuring compliance with CISA’s BOD 26-04 Prioritizing Security Updates Based on Risk (see URL in Notes) guidance and CISA’s “Forensics Triage Requirements” (see URL in Notes). Follow applicable BOD 26-04 guidance for cloud services or discontinue use of the product if mitigations are unavailable. Stakeholders are responsible for evaluating each asset's internet exposure and ensuring adherence to BOD 26-04 patching guidelines.

CVSS Score

Vector String
CVSS:3.1/AV:N/AC:L/PR:N/UI:N/S:U/C:H/I:L/A:NOpen in Calculator
Attack Vector
NETWORK
Attack Complexity
LOW
Privileges Required
NONE
User Interaction
NONE
Scope
UNCHANGED
Confidentiality Impact
HIGH
Integrity Impact
LOW
Availability Impact
NONE
Exploitability Score
3.9
Impact Score
4.2

CWEs

Affected Products

VendorProductVersion
litellmlitellm< 1.84.0

References

Weakness Type

CWE-287: Improper Authentication

Improper authentication occurs when software does not adequately prove that an actor is who they claim to be, including cases where a failed check is silently converted into success. In CVE-2026-59822 a fabricated Authorization header pushes LiteLLM onto an OAuth2 passthrough fallback in which the failure of LiteLLM key validation is replaced with an empty authentication object, so an unauthenticated caller is treated as an authenticated principal.

Learn more: CWE-287 — Improper Authentication

CWE-306: Missing Authentication for Critical Function

Missing authentication for a critical function describes the complete absence of an identity check in front of functionality that requires one. That is the practical effect on LiteLLM's MCP Streamable HTTP endpoint: once the fallback path substitutes an empty UserAPIKeyAuth(), the MCP tooling behind it is invoked with no authenticated identity at all.

Learn more: CWE-306 — Missing Authentication for Critical Function

Impact Analysis

CVE-2026-59822 is remotely exploitable over the network with low attack complexity, no authentication and no user interaction — sending one request with a fabricated Authorization header to the MCP Streamable HTTP endpoint is enough. The confidentiality impact is High: an attacker reaches MCP tooling that an AI gateway exposes, which typically includes model access, connected data sources and internal tool integrations, so prompts, responses and connector-reachable data are exposed and the organisation's paid model quota can be consumed. The integrity impact is Low and availability is unaffected, reflecting that the attacker acts through the tooling rather than gaining arbitrary control of the proxy itself. With a CVSS 3.1 base score of 8.2 (High) and an EPSS score of roughly 0.9% (57th percentile), the statistical signal is modest, but CISA's KEV listing confirms exploitation in the wild, and Wiz's AI-infrastructure honeypot research documents that exposed AI gateway components are actively probed.

Exploit Maturity

CISA has confirmed active exploitation of CVE-2026-59822 by adding it to the Known Exploited Vulnerabilities catalog with a 16 September 2026 remediation deadline. The GitHub Security Advisory, the pull request and the patch commit are all public, so the vulnerable fallback branch can be located by diffing the fix — for a logic flaw of this kind that is effectively a recipe, since exploitation requires only a crafted header rather than memory-corruption skill. Wiz's AI infrastructure honeypot research documents that exposed AI gateway and LLM proxy components attract rapid automated probing, which matches the KEV listing. The EPSS score of about 0.9% (57th percentile) understates the situation; any internet-reachable LiteLLM instance below 1.84.0 should be assumed to have been probed.

Remediation

  1. Follow the CISA KEV required action. Apply mitigations in accordance with the vendor's instructions, comply with CISA's BOD 26-04 "Prioritizing Security Updates Based on Risk" guidance and the "Forensics Triage Requirements", and discontinue use of the product if mitigations are unavailable. The KEV due date is 16 September 2026.
  2. Upgrade LiteLLM to version 1.84.0 or later. All versions before 1.84.0 are affected; the fix is documented in GHSA-7488-6r32-c95q and delivered in release v1.84.0, with the corresponding patch commit publicly available.
  3. Restrict network access to the LiteLLM proxy and its MCP endpoint. AI gateways are frequently deployed with broad reachability for convenience; place LiteLLM behind an authenticating reverse proxy or service mesh, bind the MCP Streamable HTTP endpoint to internal networks only, and drop unauthenticated traffic at the ingress layer.
  4. Rotate keys and audit MCP usage. Because an attacker could invoke tooling without a valid LiteLLM key, rotate LiteLLM virtual keys and the upstream provider API keys the proxy holds, then review request logs and provider billing for MCP tool calls and model invocations that no legitimate client initiated.
  5. Harden authentication logic long term by never allowing a failed credential check to fall through into a permissive default — a fallback path must deny, not substitute an empty identity — by requiring authentication explicitly for every critical endpoint rather than by exclusion, and by adding regression tests that assert unauthenticated requests to MCP and admin routes are rejected.

Technical Details

CVE-2026-59822 is an improper authentication flaw (CWE-287) that manifests as missing authentication for a critical function (CWE-306) in LiteLLM's MCP Streamable HTTP endpoint. When a request arrives carrying an Authorization header, LiteLLM attempts to validate it as a LiteLLM key; if that validation fails, versions before 1.84.0 take an OAuth2 passthrough fallback branch that substitutes an empty UserAPIKeyAuth() object rather than rejecting the request. Because the downstream code treats the presence of that object as evidence of successful authentication, a fabricated header value — which is trivially attacker-controlled — converts a failed check into an authenticated-looking request that proceeds to MCP tooling. The CVSS vector CVSS:3.1/AV:N/AC:L/PR:N/UI:N/S:U/C:H/I:L/A:N records a network-reachable, low-complexity attack requiring no privileges and no user interaction, with high confidentiality impact, low integrity impact and no availability impact. The fix in 1.84.0 removes the permissive fallback so that failed key validation results in denial.

Frequently Asked Questions

Is CVE-2026-59822 being actively exploited?

Yes. CISA added CVE-2026-59822 to its Known Exploited Vulnerabilities catalog with a remediation due date of 16 September 2026, confirming exploitation in the wild. The fix commit and advisory are public and the attack requires only a fabricated HTTP header, so the barrier to exploitation is very low.

What products are affected by CVE-2026-59822?

BerriAI LiteLLM is affected in all versions before 1.84.0. The vulnerable path is the MCP Streamable HTTP endpoint, so deployments that expose MCP tooling through the LiteLLM proxy are the ones at practical risk.

How do I fix CVE-2026-59822?

Upgrade LiteLLM to version 1.84.0 or later, as described in advisory GHSA-7488-6r32-c95q. Additionally restrict network access to the proxy and its MCP endpoint, rotate LiteLLM virtual keys and upstream provider API keys, and review logs and provider billing for tool calls that no legitimate client made.

How severe is CVE-2026-59822?

CVE-2026-59822 is rated High with a CVSS 3.1 base score of 8.2. It is unauthenticated and network-exploitable with low complexity and no user interaction, with high confidentiality impact and low integrity impact but no availability impact; the EPSS score sits near the 57th percentile, yet confirmed KEV-listed exploitation and a public patch make immediate upgrading necessary.

CVSS Score

8.2
HIGH(8.2)

EPSS Score

EPSS Score0.84%
EPSS Percentile58.3%

Dates

PublishedJuly 8, 2026
Last ModifiedOctober 25, 2026
StatusAnalyzed
CVSS Versionv3.1

Need Help With Vulnerability Management?

Our security experts can help you prioritize and remediate vulnerabilities effectively.