Overview
The U.S. Cybersecurity and Infrastructure Security Agency (CISA) added four high-severity vulnerabilities to its Known Exploited Vulnerabilities (KEV) catalog on November 19, 2026. Among them is CVE-2026-40217, a code execution flaw in BerriAI LiteLLM — a popular open-source LLM proxy and routing framework — directly linked to active LLMjacking operations. This marks a significant escalation in adversarial targeting of AI infrastructure, moving LLMjacking from theoretical concern to confirmed, catalogued threat.
Technical Analysis
CVE-2026-40217 — LiteLLM (CVSS 8.8): An unprotected alternate channel vulnerability allowing remote attackers to execute arbitrary code via bytecode rewriting at the /guardrails/test_custom_code URI. This endpoint, intended for testing custom guardrail logic, can be reached without authentication in misconfigured deployments, giving attackers direct code execution within the LLM serving environment. Exploitation enables adversaries to hijack LLM API credentials, redirect inference traffic, or run unauthorized model queries at the victim’s cost — the defining characteristic of LLMjacking.
CVE-2026-90436 — Apache Axis (CVSS 9.8): An unsafe reflection vulnerability enabling unauthenticated RCE. Rapid7 documented exploitation against Atlassian Confluence instances that use the Axis library for XML-RPC and SOAP APIs. Within three minutes of initial access, threat actors authenticated to MySQL using Confluence’s own credentials, exfiltrated credentials, tokens, VPN configs, and private keys. Notably, the attacker scripts were assessed as likely AI-generated, demonstrating how LLMs are being operationalised in post-exploitation toolchains.
CVE-2026-90817 — REDCap (CVSS 9.8): A code injection flaw exploitable via crafted HTTP requests targeting unintended controller routes from public survey contexts. Exploitation requires a valid public survey hash but otherwise requires no authentication.
CVE-2024-10441 — Synology DSM (CVSS 9.8): An improper output encoding vulnerability in Synology DiskStation Manager enabling RCE via unspecified vectors, linked to ransomware deployment.
Framework Mapping
- AML.T0047 (AI-Enabled Product or Service): LiteLLM is core AI serving infrastructure; its compromise directly impacts LLM availability and integrity.
- AML.T0010 (AI Supply Chain Compromise): LiteLLM’s role as a routing proxy means a single compromised instance can affect multiple downstream LLM integrations.
- AML.T0040 (AI Model Inference API Access): LLMjacking is fundamentally an unauthorized inference API access attack — hijacking paid API quotas.
- LLM05 (Supply Chain Vulnerabilities): The LiteLLM flaw exemplifies risks introduced by third-party LLM infrastructure components.
- LLM07 (Insecure Plugin Design): The exploited
/guardrails/test_custom_codeendpoint represents an insecurely designed extensibility surface.
Impact Assessment
Organizations using LiteLLM as an LLM gateway face the highest AI-specific risk: credential theft, unauthorized inference spend, and potential data leakage through hijacked model sessions. Confluence deployments with XML-RPC enabled are at critical risk of full environment compromise. The use of AI-generated attacker scripts signals a maturation of threat actor tooling, reducing the skill floor for sophisticated post-exploitation activity.
Mitigation & Recommendations
- LiteLLM: Immediately restrict or disable the
/guardrails/test_custom_codeendpoint. Apply vendor patches. Rotate all LLM API keys accessible from the LiteLLM environment. - Confluence: Disable Remote API (XML-RPC & SOAP) unless operationally required. Upgrade to Confluence 10.0+ where the feature is fully removed.
- Synology DSM: Apply the latest DSM patches. Remove unnecessary internet exposure of NAS management interfaces.
- REDCap: Apply vendor patches. Audit public survey configurations and restrict data import processing endpoints.
- General: Monitor for anomalous LLM API spend as an LLMjacking indicator. Review CISA KEV catalog for patching prioritisation.