LangChain, Langflow, LiteLLM: When AI’s Foundation Code Becomes the Attack Surface

In late March 2026 a cluster of critical incidents targeted AI infrastructure: multiple high-severity vulnerabilities in LangChain/LangGraph (path traversal, a CVSS 9.3 serialization injection dubbed "LangGrinch", and SQL injection), a Langflow unauthenticated RCE (CVE-2026-33017) weaponized within 20 hours, and a sophisticated LiteLLM supply-chain compromise that published malicious PyPI packages with credential-stealing malware after compromising a Trivy-based CI/CD action; these events demonstrate active exploitation, broad blast radius through machine identities, and the urgent need for runtime detection, credential rotation, dependency pinning, and zero-trust controls for AI frameworks.