Skip to the content
AI and Cybersecurity Learning and Practice Center Test automation Open the Learning Center

IncidentAI-enabled attacks

Symantec honeypot: an attack agent hijacks exposed AI endpoints to plan intrusions and plant miners, at the owner's expense

HighSymantec

Symantec reported on October 21, 2026 that its AI honeypot, which poses as an unprotected internet-facing AI service, captured in summer 2026 the full instructions of an attack agent whose operators wanted a real model to drive it. The request carried a jailbreak presenting the harvesting of configuration and .env files, API keys and database schemas as development work, banned refusal words such as "I can't", gave the model a persona said to run on GPT-4o, and defined about 74 offensive tools: internet-wide scanning, exploits for VS Code, Django, Jupyter, Laravel, Supabase, Next.js and Flask/Werkzeug, SSRF through HTML-to-PDF generators, cloud metadata access, database dumping and botnet control. A "Skynet" mode authorized up to 100,000 steps without human approval. A day later another request, from a rotated address, asked the model to decode a gzip and base64 shell script that installs the XMRig Monero miner with four failover pools, hides it as .trace in go2rtc directories, locks its configuration and keeps deleting local API logs. The honeypot recorded but never ran any of it, and logged more than 127,000 requests from tens of thousands of addresses. Symantec names no actor or country.

Replay the attack step by step in the Learning Center

How it happened

  1. Scan the internet for model servers, AI gateways, test systems and MCP servers that answer without authentication; most share an OpenAI-compatible interface.
  2. Send the found model a jailbreak and a set of attack tools, so the victim's model plans and drives the intrusion and the victim pays for it.
  3. Deploy a miner through a script the model is asked to decode, hidden in camera gateway directories.

Impact

Anyone exposing an AI endpoint pays for attacks on others, and the attack tools aim at stealing databases and credentials and installing miners; Symantec names no victims.

The fix

Require authentication on every AI endpoint, model-listing and health-check paths included, and do not bind model servers to all interfaces. Treat AI API keys as high-value credentials, scope and rotate them, and watch for unusual inference volume or spend. Restrict outbound traffic from AI and agent hosts, enforce IMDSv2 with a hop limit of one, and keep the targeted frameworks patched and off the internet.

Lessons

  • An open AI endpoint is not only a cost risk: attackers use it as the reasoning engine of their own operation.
  • Health checks and model lists count as endpoints; attackers probe them first.

Sources