CriticalCVE-2026-42208

    LiteLLM: SQL injection in Proxy API key verification (CVE-2026-42208)

    Published Last modified In CISA KEV since Sources: OSV, CVE.org, pwn.watch, EUVD, BSI

    What is affected, what to do

    LiteLLM is a proxy server (AI Gateway) to call LLM APIs in OpenAI (or native) format. From version 1.81.16 to before version 1.83.7, a database query used during proxy API key checks mixed the caller-supplied key value into the query text instead of passing it as a separate parameter.

    The vulnerability is being actively exploited; treat it as a priority. Update litellm to version 1.83.7, LiteLLM to version 1.83.7. Immediate actions, mitigation and detection rules are in the runbook below.

    Affected
    • BerriAI LiteLLM gte 1.81.16, < 1.83.7..*

    Not sure whether and how you are affected? Ask a question about this vulnerability

    Metrics

    9.8

    Priority

    XIEM Effective CVSS: highest available CVSS rating from NVD and vendor data.

    Source: NVD, CVSS v3

    9.8

    CVSS v3

    Severity Critical on the scale from 0 to 10.

    Source: OSV, CVE.org, pwn.watch, EUVD, BSI

    89.4 %

    EPSS score

    High — model estimates ≥ 50% chance of real-world exploitation within 30 days.

    Source: FIRST EPSS, percentile 99.78

    Show all metrics
    Severity
    Critical
    Priority (0–325)

    The priority score is a consolidated metric that condenses several individual signals into a single, action-guiding ranking: technical severity (CVSS), exploitation likelihood (EPSS), known-exploited status (KEV) and the context of the affected asset. The aim is to answer the only truly important question in vulnerability management: what first?

    Read in glossary
    245.88 Methodology
    Exploit maturity

    An exploit is code or a technique that specifically uses a concrete vulnerability to force unintended behaviour — such as executing code or escalating privileges. It is the tool that turns a theoretical flaw into a practical attack.

    Read in glossary
    actively exploitedSource: VulnCheck KEV (opens in a new tab) How we classify
    XIEM Effective CVSS

    XIEM® is the tiered managed-security operating model from NEOSEC and a registered trademark of NEOSEC GmbH. It extends classic SIEM — the central collection, normalisation and correlation of security-relevant events — with endpoint forensics, network visibility and deception technology, and consolidates them into three building tiers: Sentry, Orchestrate and Command.

    Read in glossary
    9.8 Methodology
    CVSS v3

    CVSS is the globally established, open standard for assessing the technical severity of a vulnerability on a scale from 0 to 10. It describes how easily a flaw can be exploited and what damage it can cause — but not how likely its actual exploitation is. CVSS and the complementary EPSS deliberately answer different questions.

    Read in glossary
    9.8
    CVSS v4

    CVSS is the globally established, open standard for assessing the technical severity of a vulnerability on a scale from 0 to 10. It describes how easily a flaw can be exploited and what damage it can cause — but not how likely its actual exploitation is. CVSS and the complementary EPSS deliberately answer different questions.

    Read in glossary
    9.3
    EPSS score

    The EPSS score is the absolute model output for an individual vulnerability: the estimated probability of real-world exploitation within the next 30 days, expressed as a value between 0 and 1 (i.e. 0 % to 100 %). A score of 0.42 means „the model estimates a 42 % chance of real-world exploitation within 30 days“.

    Read in glossary
    89.4 %
    High — model estimates ≥ 50% chance of real-world exploitation within 30 days.
    EPSS percentile

    The EPSS percentile is the relative position of a vulnerability within all EPSS-scored CVEs, expressed as a value between 0 and 1. A percentile of 0.90 means „this CVE has a higher EPSS score than 90 % of all other scored CVEs“. Unlike the EPSS score, it is not a probability — but a positioning within a ranking.

    Read in glossary
    99.78
    Critical — this CVE ranks in the top fifth of all CVEs scored today (rank ≥ 80%).
    KEV status

    The KEV catalogue is a list maintained by the US authority CISA of vulnerabilities for which active exploitation in the wild has been confirmed. It answers not the question „how severe?“ but „is it already being attacked?“ — thereby sharply prioritising patching.

    Read in glossary
    Actively exploited
    Published
    In CISA KEV since

    The KEV catalogue is a list maintained by the US authority CISA of vulnerabilities for which active exploitation in the wild has been confirmed. It answers not the question „how severe?“ but „is it already being attacked?“ — thereby sharply prioritising patching.

    Read in glossary
    CWE

    CWE is a catalogue of typical weakness types in software and hardware — such as flawed input validation, unsafe memory handling or poor authentication. Where a CVE names a concrete vulnerability in a concrete product, a CWE describes the underlying class of defect.

    Read in glossary
    CWE-89, CWE-89

    Response and mitigation

    Why act now?

    Prioritisation rationale

    This flaw is reachable over the network on any LLM route with no authentication or user interaction, and exploitation is trivial. One crafted header exposes the central credential store of an AI gateway. The payoff for attackers is the aggregated provider keys, which enable LLM-jacking, cost fraud and prompt/data exfiltration, plus takeover of the proxy itself. CISA does not record any known ransomware use. Exploitation likelihood is high, ranking among the more likely-to-be-exploited CVEs. Treat internet-facing instances as critical: patch and rotate secrets within 24 hours. Patch internal-only instances in the next emergency maintenance window, but still rotate their secrets.

    1. Runbook step 1

      Immediate response (0-24 h)

      • Inventory: Find every LiteLLM proxy instance, including the PyPI package litellm, ghcr.io/berriai/litellm* container images, Helm releases and developer sandboxes. Check versions with pip show litellm or the image tag. Every version from 1.81.16 up to (but excluding) 1.83.7 is affected, including 1.82.1, 1.82.3, 1.82.4, 1.82.5 and 1.82.6.
      • Patch: Upgrade to LiteLLM 1.83.7 or later, covering both container images and the PyPI package. Confirm afterwards that no stale replicas or old pods from a stalled rolling update are still serving traffic.
      • Cut exposure now: If the proxy is reachable from the internet or broad client networks (default port 4000; verify for your deployment), restrict it to known application servers via security group, firewall or ingress allowlist until the patch is confirmed.
      • Rotate secrets: Treat the proxy database as compromised on any vulnerable instance that was reachable. Rotate the following:
        • LITELLM_MASTER_KEY and all virtual keys;
        • every upstream provider key stored in the proxy (OpenAI, Azure OpenAI, Anthropic, Bedrock etc.);
        • the database credentials in DATABASE_URL. Revoke provider keys at the provider, not just in the proxy.
      • Preserve evidence: Before rebuilding, snapshot the PostgreSQL database and collect LiteLLM and reverse-proxy logs. Also pull LLM provider billing and usage data. You will need this to decide on NIS2 reporting within the 24-hour early-warning window.

    2. Runbook step 2

      Mitigation layers

      • Network segmentation: Run the proxy only in an internal application segment. Make PostgreSQL (5432) reachable solely from the proxy hosts, and restrict proxy egress to the provider endpoints you actually use.
      • WAF / reverse-proxy rule: Strictly validate the Authorization header in front of LiteLLM. Allow only Bearer sk-[A-Za-z0-9_-]+ (adapt to your key format) and reject everything else with 400. Also block SQL metacharacters and keywords in the header: ', --, ;, /*, UNION, SELECT, pg_sleep, CHR(, ||. ModSecurity example: SecRule REQUEST_HEADERS:Authorization "@rx (?i)('|--|;|union|select|pg_sleep)" "id:1042208,phase:1,deny,status:400".
      • Database least privilege: Give the proxy's DB role access to its own schema only. That means no superuser, no CREATE ROLE, and no pg_read_file or COPY ... PROGRAM. Do not share the PostgreSQL instance with other business applications, or the injection's blast radius grows with it.
      • Credential design: Scope provider keys tightly with spend caps and project-level keys. Prefer workload or managed identity over static keys where supported, so a database dump does not yield long-lived secrets.
      • Admin surface: Keep the admin UI and management routes (/key/*, /user/*, /team/*, /ui) off the public internet. Allow them only from an admin network or a VPN with MFA.

    3. Runbook step 3

      Detection rules

      • Reverse-proxy / WAF logs: Look for Authorization headers with abnormal length or SQL syntax on LLM routes (/chat/completions, /v1/chat/completions, /completions, /embeddings). Temporarily enable header logging upstream and lock down access to those logs. KQL sketch: WafLogs | where Uri has "completions" and AuthHeader matches regex @"(?i)('|--|union|select|pg_sleep)".
      • Response patterns: Bursts of 401/500 responses, or erratic response times from a single source on LLM routes, suggest blind or time-based SQLi. SPL: index=proxy uri_path="*completions" status IN (401,500) | stats count, stdev(response_time) by src_ip | where count>50.
      • PostgreSQL logs: Set log_min_error_statement = error and alert on syntax errors (syntax error at or near, unterminated quoted string) in queries against the token/verification tables. Also hunt in pg_stat_statements or pgaudit for pg_sleep, UNION SELECT and information_schema access.
      • Proxy data integrity: Alert on virtual keys, admin-role users or budget changes created outside change windows. Diff the tables against a known-good baseline.
      • Provider key abuse: Watch OpenAI, Azure, Anthropic and Bedrock for usage or cost spikes and calls from unfamiliar IPs or regions. These are the clearest signal that upstream keys have been exfiltrated.

    Detection rules

    No NEOSEC detection rule is available for this vulnerability yet.

    Classification under NIS2 and CRA

    Rule-based classification from exploitation status and severity. It does not replace a case-by-case assessment.

    NIS2

    NIS2 is EU Directive (EU) 2022/2555 to raise the level of cybersecurity across the Union. It substantially broadens obligations compared with the first NIS directive: more sectors, mandatory risk management, staged reporting duties and personal accountability of senior management.

    Read in glossary
    Actively exploited: treat the vulnerability as a priority in vulnerability handling under Art. 21(2)(e) NIS2. If it is exploited in your environment, check the reporting obligation under Art. 23 (early warning within 24 hours).
    CRA

    The CRA is EU Regulation (EU) 2024/2847 setting horizontal cybersecurity requirements for „products with digital elements“ — from industrial controllers and software to connected consumer devices. For the first time manufacturers must ensure security across the entire product lifecycle, handle vulnerabilities and provide security updates.

    Read in glossary
    Since 11.09.2026, manufacturers of affected products with digital elements must report actively exploited vulnerabilities under Art. 14 CRA, with an early warning within 24 hours via ENISA's single reporting platform.
    ISO/IEC 27001

    ISO/IEC 27001 is the internationally leading standard for an information security management system (ISMS). It describes a risk-based, process-oriented framework to systematically plan, implement, review and improve information security.

    Read in glossary
    Control A.8.8, management of technical vulnerabilities: assess relevance to your systems, plan remediation and keep evidence.

    Description

    LiteLLM is a proxy server (AI Gateway) to call LLM APIs in OpenAI (or native) format. From version 1.81.16 to before version 1.83.7, a database query used during proxy API key checks mixed the caller-supplied key value into the query text instead of passing it as a separate parameter. An unauthenticated attacker could send a specially crafted Authorization header to any LLM API route (for example POST /chat/completions) and reach this query through the proxy's error-handling path. An attacker could read data from the proxy's database and may be able to modify it, leading to unauthorised access to the proxy and the credentials it manages. This issue has been patched in version 1.83.7.

    Source: OSV (Open Source Vulnerabilities), NVD (NIST), pwn.watch, CISA KEV Catalog, cvelistv5, ENISA EUVD, CVE.org (MITRE), BSI WID Portal, BSI CSAF

    Affected products

    The following versions are affected. Older releases of the same product line are also vulnerable unless stated otherwise.

    Affected products
    Affected versionsFixed in
    BerriAIlitellm
    • ≥ 1.81.16, < 1.83.7
    • —
    litellmlitellm
    • 1.81.16 – < 1.83.7
    • 1.83.7
    Open SourceLiteLLM
    • < 1.83.7
    • 1.83.7
    pypilitellm
    • 1.82.4
    • 1.83.1
    • 1.82.0
    • 1.83.0
    • 1.81.16
    • 1.82.6
    • 1.83.4
    • 1.82.2
    • 1.83.2
    • 1.83.5
    • 1.82.1
    • 1.82.3
    • 1.82.5
    • 1.83.6
    • 1.83.3
    • —
    • —
    • —
    • —
    • —
    • —
    • —
    • —
    • —
    • —
    • —
    • —
    • —
    • —
    • —

    Vendors and products

    Weakness classes (CWE)

    • CWE-89Improper Neutralization of Special Elements used in an SQL Command ('SQL Injection')

      Base

      The product constructs all or part of an SQL command using externally-influenced input from an upstream component, but it does not neutralize or incorrectly neutralizes special elements that could modify the intended SQL command when it is sent to a downstream component. Without sufficient removal or quoting of SQL syntax in user-controllable inputs, the generated SQL query can cause those inputs to be interpreted as SQL instead of ordinary user data.

      Entry at MITRE CWE-89

    • CWE-89Improper Neutralization of Special Elements used in an SQL Command ('SQL Injection')

      Base

      The product constructs all or part of an SQL command using externally-influenced input from an upstream component, but it does not neutralize or incorrectly neutralizes special elements that could modify the intended SQL command when it is sent to a downstream component. Without sufficient removal or quoting of SQL syntax in user-controllable inputs, the generated SQL query can cause those inputs to be interpreted as SQL instead of ordinary user data.

      Entry at MITRE CWE-89

    Public exploit references

    Publicly reported proof-of-concepts, exploit modules and detection templates for this vulnerability. NEOSEC links to the original sources only and does not provide any exploit code.

    Public exploit references
    CISA KEV CatalogCISA Known Exploited Vulnerability (opens in a new tab)cisa.govactively exploited
    PoC-in-GitHub (nomi-sec)CVE-2026-42208-LAB (opens in a new tab)github.comPoC (publicly reported)reported, not verified
    PoC-in-GitHub (nomi-sec)litellm-scanner (opens in a new tab)github.comPoC (publicly reported)reported, not verified

    References & sources

    • github.comBerriAI/litellm/security/advisories/GHSA-r75f-5x8p-qvmcweb, x_refsource_CONFIRM, pwn-watch, Mitigation, Patch, Vendor advisory
    • NVD (NIST)NVD: CVE-2026-42208advisory, pwn-watch
    • github.comBerriAI/litellmpackage
    • github.comBerriAI/litellm/releases/tag/v1.83.7-stableweb, x_refsource_MISC, Product, Release notes
    • cisa.govknown-exploited-vulnerabilities-catalog?field_cve=CVE-2026-42208web, government-resource, US government resource
    • pypi.orgproject/litellmpackage
    • GitHub Security Advisoriesadvisories/GHSA-r75f-5x8p-qvmcadvisory
    • access.redhat.comsecurity/cve/CVE-2026-42208Vulnerability database, x_refsource_REDHAT, Third-party advisory
    • bugzilla.redhat.comRHBZ#2463965Issue tracking, x_refsource_REDHAT, Mitigation, Third-party advisory
    • security.access.redhat.comdata/csaf/v2/vex/2026/cve-2026-42208.jsonx_sadp-csaf-vex, Third-party advisory
    • github.compypa/advisory-database/tree/main/vulns/litellm/PYSEC-2026-391.y…web
    • pwn.watchpwn.watch: BerriAI LiteLLM: BerriAI LiteLLM SQL Injection Vulnerabilitypwn-watch
    • cisa.govCISA Known Exploited Vulnerabilities Catalogpwn-watch

    Changes to this CVE

    Who changed this CVE and when: NVD (analysis, affected products, CVSS), the CVE publisher (CNA) and additional providers such as CISA (ADP).

    1. NVD analysis modifiedNVD

      Show 3 original changes
      Individual changes of this event as published by NVD
      TypeActionBeforeAfter
      Reference typeadded
      —
      redhat-SADP: https://access.redhat.com/security/cve/CVE-2026-42208 Types: Third Party Advisory
      Reference typeadded
      —
      redhat-SADP: https://bugzilla.redhat.com/show_bug.cgi?id=2463965 Types: Mitigation, Third Party Advisory
      Reference typeadded
      —
      redhat-SADP: https://security.access.redhat.com/data/csaf/v2/vex/2026/cve-2026-42208.json Types: Third Party Advisory

      Source identifier: nvd@nist.gov

    One NVD translation into another language is hidden.

    Linked advisories

    Embed badge

    Show the severity and KEV status of this vulnerability on your site – with a link back here. The image is updated hourly.

    Badge preview for CVE-2026-42208
    <a href="https://intel.neosec.eu/en/cve/CVE-2026-42208"><img src="https://intel.neosec.eu/badge/cve/CVE-2026-42208.svg?lang=en" alt="CVE-2026-42208 – NEOSEC Intel" height="20"></a>