CISA flags active LiteLLM gateway exploit
CISA added a critical sandbox-escape vulnerability in LiteLLM (CVE-2026-40217) to its Known Exploited Vulnerabilities catalog following active in-the-wild exploitation. Attackers leverage the proxy flaw in LLMjacking campaigns to extract enterprise API keys and rack up catastrophic model consumption charges.
AI gateways are becoming prime targets because they store the master API keys to an organization's entire model fleet. Compromising a proxy turns infrastructure management into an all-access pass for automated LLMjacking and credential theft.
- –Sandbox escape via custom guardrails: Tracked as CVE-2026-40217 (CVSS 8.8), the flaw in /guardrails/test_custom_code allowed remote code execution via bytecode rewriting that bypassed a custom Python sandbox.
- –Root privileges in container defaults: Standard LiteLLM container deployments ran as root, allowing attackers who breached the proxy to gain immediate system access and dump plaintext API credentials.
- –LLMjacking monetization: Threat actors are targeting AI gateways to siphon inference compute, reselling access via underground LLM mixers and running up six-figure cloud consumption bills.
- –Patch lag proves costly: BerriAI addressed the vulnerability in versions 1.83.10 and 1.83.11 back in April, but exposed, unpatched gateways left enterprises vulnerable months later.
- –Urgent remediation required: Teams operating LiteLLM instances must upgrade past 1.83.11 immediately, restrict admin endpoints behind VPNs or reverse proxies, and implement hard spend caps on provider accounts.
DISCOVERED
13m ago
2026-11-20
PUBLISHED
30m ago
2026-11-20
RELEVANCE
AUTHOR
wifeydoteth