Threat researchTracked by EncryptLayer intelligence

Sekoia Honeypot Captures an Attack on GenAI Infrastructure

Published by
sekoia.com
Published
23 Oct 2026
Indicators
none submitted

The report

Sekoia’s TDR team ran a decoy Flowise server to study attacks on GenAI platforms. One threat actor stole its credentials and tried to misuse its AI agent.

sekoia.com

Read the full analysis: Sekoia Honeypot Captures an Attack on GenAI Infrastructure (sekoia.com).

What the report says

Sekoia’s TDR team reports that a decoy Flowise server it ran as a honeypot was attacked in late August 2026 by a likely Chinese speaking threat actor, who stole its credentials and tried to misuse its AI agent. The report says the attacker mainly abused legitimate Flowise functionality rather than software vulnerabilities, and tried to turn the LLM agent into an offensive analysis assistant against external targets. The authors say GenAI orchestration platforms are becoming high value targets because they hold API keys, enterprise data and access to tools.

  • The report says an unauthenticated honeypot simulating Flowise version 2.2.8 was subjected to reconnaissance and attacks in late August 2026, during which the attacker exfiltrated credentials.
  • According to the report, the attacker submitted real world reconnaissance data on Nightingale/Redis infrastructure to the AI agent to generate custom exploitation playbooks and target analysis.
  • The authors attribute the activity, based on linguistic indicators and tradecraft, to a likely Chinese speaking actor targeting unregulated online gambling (iGaming) infrastructure, particularly in Southeast Asia. They describe it as more consistent with opportunistic cybercrime than with a bespoke APT operation.
  • The report describes automated reconnaissance followed by manual interaction. A first phase of approximately 3 minutes identified the service and stole secrets. Just over two hours of inactivity followed, then systematic scanning of API endpoints, re-exfiltration of credentials, inference probing with a basic prompt, and attempts to modify and create chatflows.
  • The report says the honeypot monitors exploitation attempts against CVE-2025-59528 (CustomMCP remote code execution), CVE-2026-41264 (CSV Agent remote code execution) and flaws similar to CVE-2024-31621 (authentication bypass and credential disclosure).
  • The report says the honeypot uses canary tokens, including credentials for a simulated production PostgreSQL database used for RAG, but cannot detect use of fake third party tokens such as OpenAI API keys.

Named in the report: Flowise, Nightingale, Redis.

CVEs: CVE-2024-31621, CVE-2025-59528, CVE-2026-41264.

This digest was written from the report text itself by a language model and checked by script: every number and CVE id in it appears in the report. The report is the source; read it for the full detail.

Indicators

The report is tracked for context; no indicator from it met the bar for blocking (for example, it only names shared hosting).

How this is tracked

EncryptLayer intelligence, A.R.P. Syndicate’s open threat feed, reviews each indicator against the published report before adding it (#25). This page is generated from that record and the report itself; the indicators are never rewritten.