2026-06-16 · view entry permalink →
Obsidian Security: a three-CVE chain turns any LiteLLM user into root on the AI gateway
Obsidian Security published a privilege-escalation-to-RCE chain in LiteLLM (BerriAI), the widely self-hosted AI gateway that proxies 100+ LLM providers behind one OpenAI-compatible API (Obsidian Security, 2026-06-15; The Hacker News, 2026-06-15). The chain: CVE-2026-47101 (authorization bypass), the key-generation endpoint accepts a caller-supplied allowed_routes without checking the caller's role, so an internal_user can mint a key reaching admin routes; CVE-2026-47102 (privilege escalation); /user/update lacks field-level authorization, letting any authenticated user set their own user_role to proxy_admin; CVE-2026-40217 (RCE); the Custom Code Guardrails feature runs attacker-supplied Python via exec() with __builtins__ available, giving arbitrary code execution. VulnCheck scores CVE-2026-47102 at CVSS 8.8 (3.1), and Obsidian rates the chained impact CVSS 9.9; chained, a default low-privilege account reaches the master key, the salt key decrypting stored secrets, the database URL and every configured provider API key, and can rewrite responses delivered to downstream AI agents ("man-in-the-gateway"). Fixed in v1.83.14-stable, but Obsidian reports broad under-deployment of the fix. Mapped to T1078, T1548 and T1059.006.
Why it matters to us: Swiss/EU public-sector and research bodies increasingly centralise AI workflows on a gateway proxy; a compromised LiteLLM is both a credential-theft and an agent-manipulation vector. Pin LiteLLM to ≥1.83.14, keep admin endpoints off the internet, store provider keys in a secrets manager, and rotate all provider keys if any pre-1.83.14 instance was reachable by untrusted users.