CVE-2026-19297

CVE-2026-19297 is a missing rate-limit on Langflow's authentication endpoint, a vulnerability class so well-established it barely warrants a CVE in 2026. The CVSS 9.1 reflects theoretical maximum impact: an attacker can brute-force credentials indefinitely. The EPSS of 0.00418 says exploitation is statistically unlikely. That gap is real, but it's not the most important thing to understand about this vulnerability. Langflow is not a traditional web application. It's an LLM orchestration platform that authenticates against GPT-4, Claude, and local model endpoints; stores conversation histories; manages workflow configurations; and often has network access to internal data sources. A missing rate limit here doesn't mean 'attacker guesses passwords.' It means automated credential stuffing at scale against a system that holds API keys for expensive AI resources, prompt templates, data pipeline configurations, and cached outputs. The blast radius extends through every API key Langflow maintains, every conversation it proxies, every workflow it executes on behalf of compromised users. This context fundamentally changes the severity calculus in ways neither CVSS nor EPSS capture. CVSS measures a standalone authentication flaw in isolation. EPSS measures historical exploitation probability for this vulnerability class. Neither accounts for what a compromised Langflow session actually gives an attacker: an operational AI gateway with compute budget and data access. For defenders, the priority is straightforward but not simple: patching the rate limit is necessary but insufficient. Audit what credentials Langflow holds—API keys for model providers, database connections, internal service access. Treat Langflow as a high-value target because it is one. The vulnerability class is solved at the code level; the deployment context is what makes this critical. Assume that if an attacker gains access, they inherit everything Langflow can reach, not just a login session.

Reviewed through automated stages and approved by a human before publication.