Era’s researchers observed attackers sending 4,972 messages that all said “hello” to deliberately exposed Flowise installations. The replies contained host information, credential searches and cloud-metadata results. Commands traveled in request-time configuration, leaving the saved workflow looking harmless.

The security vendor’s November 3 report calls the framework Greeter. Its observations came from seven honeypots across four regions, monitored from July to October 2026. Honeypots are systems deliberately exposed to observe attacks.

Evidence status: source-observed exploitation of instrumented honeypots. Era supplied fake cloud credentials. Its report does not establish theft from real customer cloud accounts, a worldwide victim count or successful account-wide takeover.

The executable override matters more than the greeting

Flowise lets developers build applications around language models. Its prediction API accepts runtime configuration through overrideConfig. On the vulnerable setup, a caller could replace the code inside a Custom JavaScript Function node for an individual request.

Era ran Flowise 1.8.2 with application authentication deliberately unset. The attack used the code override to execute commands and receive their output. Reviewing only stored chatflow definitions or visible user messages would miss the changing instructions.

The underlying issue predates this campaign. Flowise published GHSA-5cph-wvm9-45gj on November 21, 2024, listing versions below 2.1.4 as affected and 2.1.4 onward as patched for that advisory. The later CVE-2024-58351 record describes the same configuration-injection problem.

That repair boundary applies to this particular flaw. It is not a recommendation to deploy an old 2.1.4 release today. Era discusses other, later Flowise vulnerabilities separately.

The relevant exposure combines reachable application APIs, vulnerable executable overrides and permissions available to the Flowise process. Era’s unauthenticated honeypots demonstrate that path; they do not establish identical exposure for every authenticated or differently configured deployment.

Credential discovery is observed; broader takeover is conditional

Era reconstructed host reconnaissance, searches for configuration and SSH credentials, cloud-metadata probes and attempts to establish persistence. The report says the persistence deployment failed in the observed case. Checks for memory-based execution showed capability, rather than proving that a payload ran that way.

The observed process already had root privileges, and the environment exposed powerful host interfaces. Those choices affect what the attacker could discover and attempt. They should not be treated as defaults for every Flowise installation.

A compromised workload may obtain fresh credentials from its cloud identity. What follows depends on that identity’s permissions and trust relationships. Era’s discussion of role escalation and administrative takeover describes possible next stages, not completed outcomes demonstrated in a production cloud account.

Chinese-language labels describe the framework’s output. They do not establish the operator’s nationality. Likewise, regularly spaced requests suggest automation without establishing that an AI agent controlled the attack.

Contain both the application and its identity

For operators of Flowise or similar workflow services:

  • Update through the project’s supported release process and authenticate exposed APIs. Restrict which runtime settings untrusted callers may change, especially executable code.
  • Investigate request-time changes as well as saved workflows. Preserve relevant application and host evidence. Repeated greetings alone are weak evidence; pair them with unexpected execution results and configuration overrides.
  • Run with narrowly scoped privileges. Avoid root, unnecessary Linux capabilities and access to host administration interfaces such as the Docker socket.
  • Limit the workload’s cloud identity. Restrict resource access, role passing, role assumption and permission changes to the application’s actual needs.
  • If compromise is suspected, address exposed credentials and the ability to mint replacements. Patching the application cannot recall copied secrets or terminate every identity-based foothold.

Era recommends requiring AWS’s newer metadata protocol, IMDSv2, where applicable. It also explains its limit: code already running inside a workload may request the metadata session token itself. Metadata hardening complements least privilege rather than containing arbitrary code execution on its own.

Caveat: Era sells cloud-security services; its honeypot observations and version-based exposure estimates have not been independently reproduced in the reviewed sources.