Vulnerability  ·  2026-08-15

Flowise AI agent builder — regex-based Python sandbox validator bypass across CSV/Airtable Agent nodes (cluster of 4 CVEs)

VulnerabilityHigh impactGlobalCVE-2026-73485
NVD published a cluster of Critical Flowise CVEs on 2026-08-13: CVE-2026-73485/73486/73487 (CVSS 9.0) describe regex-blocklist bypasses in Airtable/CSV Agent Python code validators enabling unauthenticated code injection via prompt injection, and CVE-2026-73601 (CVSS 9.0) describes RCE in the Custom MCP node via stdio protocol environment manipulation. All are fixed in Flowise 3.1.3.
Flowise is one of the most widely deployed open-source no-code LLM agent builders; this is a direct, unauthenticated prompt-injection-to-RCE chain — attackers can exfiltrate connected datasets, pivot via SSRF, or fully compromise the Flowise server merely by sending a crafted prompt.
A prompt injection sent to a chatflow using the Airtable/CSV Agent node causes the LLM to emit a Python script that bypasses the static regex-based pythonCodeValidator blocklist via import-aliasing/obfuscation, then executes unsandboxed in Pyodide; a related Custom MCP node flaw (CVE-2026-73601, CVSS 9.0) allows RCE via stdio protocol environment-variable manipulation
Flowise < 3.1.3 (Airtable Agent, CSV Agent, and Custom MCP nodes)
Upgrade to Flowise 3.1.3 or later; review chatflows for CSV/Airtable Agent and Custom MCP nodes
GitHub Advisory GHSA-c5hr-rc98-xp3gGitHub Advisory GHSA-4878-cqgq-j53vGitHub Advisory GHSA-g98q-rm45-q9h8CISA Vulnerability Summary SB26-222
See this in the live feed Explore related AI security and governance findings — updated every morning.
Open the feed →