Vulnerability  ·  2026-06-18

LiteLLM Low-Privilege → Admin → RCE Privilege Escalation Chain (CVSS 9.9, Obsidian Security)

VulnerabilityHigh impactGlobalCVE-2026-47101
Obsidian Security disclosed a CVSS 9.9 three-vulnerability chain in LiteLLM on June 11, 2026, responsibly reported to BerriAI in February 2026 and fully patched in v1.83.14-stable. CVE-2026-47101 allows any internal user to mint an API key with wildcard route access; CVE-2026-47102 allows self-promotion to proxy_admin via an unprotected user_role field; CVE-2026-40217 allows RCE through the Custom Code Guardrail's unsandboxed exec() call. Additionally, Obsidian demonstrated a novel response-injection attack: a compromised proxy can silently rewrite model responses in transit using built-in callbacks, injecting malicious tool calls that never reach the model, bypassing prompt injection defenses entirely.
This chain demonstrates that AI gateways, long treated as passive middleware, are now first-class attack targets. Beyond credential theft, the response-injection technique is qualitatively new: it does not manipulate the LLM — it intercepts the wire between the model and the agent, converting the gateway into an agent hijacking device. Every agent routing through a compromised LiteLLM proxy can be silently redirected. The blast radius includes all downstream AI agent workflows, CI/CD pipelines using AI-assisted code review, and MCP-connected tools.
Three-step chain from a default low-privilege internal_user account: (1) CVE-2026-47101 — create an API key with allowed_routes:["/*"] to bypass route-level RBAC; (2) CVE-2026-47102 — POST user_role:"proxy_admin" to /user/update to self-promote to full admin; (3) CVE-2026-40217 — use the Custom Code Guardrail exec() endpoint (no builtins filtering) to pop a reverse shell. Obsidian also demonstrated response-injection against Claude Code routed through the compromised proxy — injecting a malicious tool call that delivered a reverse shell on a developer machine from a single 'hello' prompt.
BerriAI LiteLLM < 1.83.14-stable (CVE-2026-47101, CVE-2026-47102, CVE-2026-40217)
Upgrade to LiteLLM ≥1.83.14-stable (all three CVEs patched). Published by Obsidian Security on June 11, 2026; reported to BerriAI in February 2026.
Sources
Obsidian Security — Breaking LiteLLM (primary advisory, verified full text)latesthackingnews.com — LiteLLM Vulnerability Chain analysis (verified full text)The Hacker News — LiteLLM Vulnerability Chain Lets Low-Privilege Users Take Over AI Gateway Servers
See this in the live feed Explore related AI security and governance findings — updated every morning.
Open the feed →