AIDEFEND in Action AIA-0145 Incident

Flowise Honeypot: Unauthenticated API Abuse, Beacon Setup, and Agent Probing

Sekoia recorded six phases of real attacker traffic against an internet-facing decoy that emulated Flowise 2.2.8, a platform for building AI workflows. The exposed management API let the attacker retrieve planted credentials and add a workflow configured to call back. The attacker then tried to override prompts, probe session memory and ask the simulated agent about real third-party infrastructure. The decoy returned simulated responses, so the requests do not establish real command execution, database access or a Nightingale compromise.

Credential TheftPrompt InjectionAI Deployment SecurityAI InfrastructureRAG Security
10 applicable AIDEFEND defenses
Source: A honeypot's tale: Deconstructing an attack on GenAI infrastructure 
Authors: Jérémy Scion, Sekoia TDR Team
Original article: Oct 19, 2026

Threat Analysis

  • The decoy exposed management routes on purpose. Sekoia built a small API emulator for Flowise 2.2.8 with fake streaming output and state separated by source IP. The attack did not depend on an exploited CVE. In the first phase, lasting under three minutes, the attacker fingerprinted the service, enumerated workflows and requested individual credential records. OpenAI, Anthropic and PostgreSQL credentials returned by the decoy were planted for the study.
  • The workflow revealed an intended callback. After a roughly two-hour pause, the attacker swept 14 management routes and probed prediction. They tried different methods to rename a workflow to Customer Support v2 and submitted a new System Status flow containing an outbound HTTP Request node. The setup pointed to an attacker-controlled destination. The report establishes the workflow configuration, without establishing that a real job made the call.
  • The prediction requests reveal what the attacker wanted. The attacker requested system commands and prompt disclosure, tested memory across turns and sessions, and supplied an overrideConfig.systemMessage field to try to change system instructions. They then presented a supposed security assessment to ask for finance retrieval-augmented generation (RAG, model answers grounded in retrieved data) and network reconnaissance. Both internal targets and responses were simulated.
  • The external Nightingale target was real. The attacker asked for a Chinese-language analysis of Nightingale v6, an infrastructure monitoring system, and included names of Redis entries from a real target. Sekoia found the same names in an exposed Redis deployment while checking public Shodan data. It inferred that the attacker had probably identified the target through Shodan or a similar source. Phase six then polled the decoy variables and chatflows at roughly ten-minute intervals.

Applicable AIDEFEND Defenses (10)

AID-H-004.002
Service & API Authentication
Very High
Require authenticated, scoped API credentials and administrative authentication across every Flowise management and workflow route (/api/v1/credentials, /api/v1/chatflows, /api/v1/variables, /api/v1/node-configs, /api/v1/apikey, and /api/v1/users), and place the management plane behind an authenticating gateway or private network. Because the entire Phases 1 to 4 intrusion relied on unauthenticated REST access rather than a software exploit, enforcing authentication at the API boundary independently blocks anonymous enumeration, credential theft, and workflow creation.
AID-E-001.001
Root & Long-Lived Credential Object Eviction
Very High
For any real Flowise deployment exposed without authentication, treat all secrets stored in /api/v1/credentials and /api/v1/variables (including provider API keys and database passwords) that incident evidence brings into scope as potentially exposed. Close the unauthorized route, then rotate or revoke the exact credentials with the responsible providers or database administrators.
AID-H-002.002
Inference-Time Prompt & Input Validation
High
On a production prediction route, validate the JSON against a closed allowlist before forwarding it to Flowise. Reject untrusted overrides of server-owned system instructions or tool configuration. Keep session binding separate and verify that authenticated callers cannot reach a less restrictive direct prediction route.
AID-H-017.002
Least-Privilege Tool Architecture
High
Remove generic ambient-power tools such as unrestricted HTTP Request nodes, shell/command executors, and raw database query tools from externally reachable chatflows. Expose only single-purpose, strongly typed tools backed by least-privilege read-only credentials scoped to the exact data a public assistant is allowed to return.
AID-I-001.004
Sandbox Network Egress Restrictions
High
Enforce an external network policy that allows the Flowise runtime to connect only to approved external services. Apply it outside the process so an HTTP workflow node cannot send callbacks to an unapproved host.
AID-I-002.001
Internal AI Network Segmentation
High
Segment production Flowise workers from unrelated internal hosts. On Kubernetes, use a policy-enforcing CNI and allow those workers to reach only their approved internal services and ports. Even if a workflow accepts a request to probe the internal network, the worker cannot connect to hosts outside that list.
AID-I-004.001
Runtime Context Isolation & Hygiene
High
Bind volatile conversation memory to a server-minted, authenticated session identifier with strict TTL and size limits rather than accepting caller-chosen sessionId strings (such as tame-session-001), preventing an attacker from reading another user’s conversation history or pre-seeding shared session memory.
AID-DV-001
Honeypot AI Services & Decoy Models/APIs
High
An authorized threat-research or defense team can deploy an isolated, manifest-defined decoy AI service (like Sekoia’s Flowise 2.2.8 REST emulator) seeded with synthetic node configurations, chatflows, canary credential UUIDs, and simulated event streams with zero connection to real LLMs, OS shells, or production networks. This safely captures attacker tooling, credential harvesting, beacon infrastructure (154.217.251.226:19999), and downstream victim intelligence (Nightingale v6 keys).
AID-E-003.003
Confirmed Malicious Code & Persistence Eviction
High
For a production compromise, isolate the affected workflow first. Compare workflow definitions with the approved baseline, preserve evidence, then remove the confirmed malicious workflow and unauthorized edits from the authoritative configuration and running deployment. Verify the affected workflow cannot be loaded or invoked.
AID-I-003.002
Identity, Session & API Rate-Limit Throttling
Medium
At the gateway in front of exposed API routes, enforce base rate limits by verified caller identity or anonymous tier. Use trusted edge abuse signals, such as source reputation or repeated polling, to impose stricter temporary limits when supported. This reduces the rate of scanning and inference misuse when requests exceed the allowance or meet an abuse restriction. Low-frequency requests can remain within the allowance.

What Defenders Should Do Now

  • Put administrative routes behind authenticated access and enforce separate permissions for listing, reading credential records and changing workflows. If inference is public, use a separate ingress with a closed schema and only narrowly scoped tools.
  • Keep secrets at the provider or credential broker and turn off arbitrary networked workflow features in public-facing flows. Verify that the running service cannot bypass its network restrictions.
  • Investigate actual exposure by correlating API calls, workflow changes, upstream credential use and outbound telemetry. If a real deployment was affected, remove unauthorized workflow definitions and invalidate exposed credentials at their issuers.

Conclusion

The decoy shows which parts of a low-code AI platform an attacker probes first. AIDEFEND  maps the production defenses to authenticated management ingress, controlled inference inputs and least-privilege workflow actions. Keeping those boundaries separate makes it possible to offer inference without exposing workflow administration or backend credentials.

Last updated: