LiteLLM CVE-2026-42271 MCP stdio command injection
Summary
CISA added CVE-2026-42271 to the Known Exploited Vulnerabilities catalog on June 8, 2026, citing active exploitation of a BerriAI LiteLLM command-injection vulnerability.
GitHub's advisory describes the bug as authenticated command execution through two LiteLLM MCP REST preview endpoints: POST /mcp-rest/test/connection and POST /mcp-rest/test/tools/list. In affected versions, those endpoints accepted full MCP server configuration from the request body, including stdio command, args, and env fields. When the proxy tried to test a stdio server, it spawned the supplied command as a subprocess on the LiteLLM proxy host.
The vulnerable range is LiteLLM 1.74.2 through versions before 1.83.7. The patched version is 1.83.7.
Tags
- ops
- operations
- active exploitation
- CISA KEV
- AI gateway
- LiteLLM
- MCP
- Model Context Protocol
- stdio
- command injection
- RCE
- CVE-2026-42271
- CWE-78
- CWE-77
Why this matters
- LiteLLM is commonly deployed as an AI gateway / proxy in front of model providers; compromise can expose model credentials, API keys, chat data, and adjacent internal services reachable from the proxy host.
- The endpoints were gated by a valid proxy API key, but GitHub says there was no role check: any authenticated user, including low-privilege internal-user key holders, could trigger host command execution.
- This is a concrete exploited instance of the broader MCP stdio boundary: stdio transport is meant to execute commands, so applications must not let untrusted users provide raw
command/argsvalues. - CISA's KEV due date for covered U.S. federal civilian agencies is June 22, 2026; private defenders should treat internet-exposed or broadly shared LiteLLM proxies as urgent.
Reported vulnerable path
- A user with any valid proxy API key sends a request to one of the MCP preview endpoints:
POST /mcp-rest/test/connectionPOST /mcp-rest/test/tools/list- The request body supplies a stdio MCP server configuration, including attacker-controlled
command,args, orenvfields. - LiteLLM attempts to preview or list tools for the configured server.
- The proxy host spawns the supplied command with the privileges of the LiteLLM proxy process.
Defender actions
- Upgrade LiteLLM to 1.83.7 or later.
- If immediate upgrade is not possible, restrict access to the LiteLLM proxy and especially the MCP REST preview endpoints; treat any low-privilege internal-user key as sufficient for exploitation in affected versions.
- Review LiteLLM logs and host telemetry for unexpected calls to
/mcp-rest/test/connectionor/mcp-rest/test/tools/list, followed by shell, Python, Node, curl, wget, package-manager, cloud-CLI, or archive-tool subprocesses from the LiteLLM process. - Rotate LiteLLM proxy keys and downstream model-provider / cloud credentials if exploitation is suspected.
- Run LiteLLM with least privilege: isolated service account, minimal filesystem mounts, restricted egress, no ambient cloud metadata access, and no unnecessary secret-bearing environment variables.
- Treat MCP stdio configuration as code execution. Prefer allow-listed MCP server profiles rather than letting users submit arbitrary command lines.
Related pages
Sources
- CISA KEV entry: https://www.cisa.gov/sites/default/files/feeds/known_exploited_vulnerabilities.json
- GitHub advisory GHSA-v4p8-mg3p-g94g: https://github.com/BerriAI/litellm/security/advisories/GHSA-v4p8-mg3p-g94g
- LiteLLM v1.83.7 stable release: https://github.com/BerriAI/litellm/releases/tag/v1.83.7-stable
- NVD CVE-2026-42271: https://nvd.nist.gov/vuln/detail/CVE-2026-42271