Observed incident

Flowise honeypot records credential harvesting and agent-abuse attempts

Reported

Published here

What happened

Sekoia observed an attacker harvesting planted credentials from an unauthenticated Flowise emulation, attempting workflow changes and jailbreaks, and seeking attack advice about gambling infrastructure.

Parties and reporting

AI tools and systems: Flowise 2.2.8 API emulation, Simulated LLM and RAG workflows, Nightingale / Redis reconnaissance data

Outcome

The decoy ran no real language model or system code. Successful compromise of the external targets was not established.

Operating lesson

Authenticate orchestration interfaces, limit stored credentials and tool permissions, and monitor unexpected workflow changes and outbound requests.

Evidence and limits

Observed attack attempts against a decoy in late August, privately reported September 7 and publicly described October 19. This does not establish a real Flowise deployment breach or successful LLM misuse. Attribution remains unknown.

Sources