Unauthenticated remote code execution in Langflow public flow builds
CVE-2026-33017 is a code-injection vulnerability affecting Langflow versions earlier than 1.9.0. The unauthenticated public flow build endpoint accepts an optional data parameter that substitutes attacker-controlled flow definitions for definitions persisted in the database. Python code embedded in those definitions reaches exec() without sandboxing, including through prepare_global_scope(), allowing arbitrary code execution with the Langflow process's privileges. Although the endpoint verifies that the requested flow exists and is public, it incorrectly trusts executable flow data supplied by the caller. This vulnerability is distinct from CVE-2025-3248, which concerned authentication on the separate code-validation endpoint. Langflow 1.9.0 fixes the issue.
What it means and what to do now: patch path, interim mitigations, and the assume-compromise checklist.
What an attacker gets, and what they’ve been doing with it.
Successful exploitation compromises the confidentiality, integrity, and availability of resources accessible to the Langflow process. Attackers can execute operating-system commands, read or modify files, access workflow and database data, and extract environment variables, AI-provider API keys, cloud credentials, and other application secrets. Exploitation has been observed delivering cryptocurrency-mining malware with persistence and SSH propagation capabilities. Further host compromise, security-control interference, and lateral movement depend on the service account's privileges, available credentials, and network reachability; root execution is not an inherent property of the vulnerability.
If you can’t patch tonight, do this now.
Until patched, block the public flow build endpoint at a reverse proxy or disable public flow building. A code-level workaround is to remove or ignore the optional data parameter on the unauthenticated endpoint and load flow definitions exclusively from the database. Restrict Langflow access to trusted networks or an authenticated gateway. Disable AUTO_LOGIN to prevent anonymous acquisition of administrative tokens and creation of prerequisite public flows; this alone does not prevent exploitation when an attacker already knows an existing public flow identifier. Run Langflow as a dedicated non-root account with least privilege and operating-system confinement to reduce post-exploitation impact.
Patch, then assume compromise.
Upgrade Langflow to version 1.9.0 or later. Ensure unauthenticated public flow builds use only persisted database-backed flow definitions rather than caller-supplied executable data. Investigate previously exposed vulnerable deployments for unauthorized execution, modified flows, persistence, and credential access. Rotate potentially exposed application, cloud, database, AI-provider, and SSH credentials, and rebuild confirmed compromised systems from known-good images or backups.
31 exploits validated by Mallory's analysis. Hidden: 4 rejected by Mallory's analysis as not a working exploit; 2 not yet analyzed, such as Exploit-DB entries. Showing the 25 most capable, earliest first.
Products and vendors Mallory has correlated with this vulnerability. Open in Mallory to drill down to specific CPE configurations and version ranges.
274 sources tracked across advisories, community write-ups, and news. New activity surfaces here as Mallory finds it.