Skip to content

Response, New York. Noah Park: Check Point Patches Actively Exploited Management Server Path Traversal. Detection, Dallas. Ava Okello: Report Finds Nearly Half of Deployed Detections Need Attention Before SOCs Can Trust Them.

SOCtember

Always First. Fast SOC News.

Symantec hero image of an amber light trail moving through a field of binary digits on a dark blue background.
Photo illustration: Symantec / Broadcom, header image for its Skynet LLMjacking research, Oct 21, 2026 (fair use).

Detection

Symantec Says an Autonomous Skynet Agent Hijacked Exposed AI Models to Scan, Exploit Web Frameworks, Steal Credentials and Mine Monero for Up to 100,000 Steps

Symantec's AI honeypot captured the full working prompt in the summer of 2026. The request wrapped a jailbreak around roughly 74 offensive tools, authorized unattended tool calls and pointed a later payload at go2rtc camera directories. The hashes, mining pools and what to lock down on every internet-facing model endpoint.

Ava Okello, DetectionLondon6 min read

LONDON - Attackers have built an autonomous offensive agent that hunts the internet for exposed large language model endpoints and hijacks them to run its own attacks, according to research Symantec published Wednesday.

Symantec says its AI honeypot, a system that impersonates an unprotected internet-facing AI service, captured the agent's full instruction set in the summer of 2026. Because the operators sent the entire working prompt to what they believed was a real model, the honeypot recorded an unusually complete view of how the agent is built. The request was not aimed at subverting the honeypot itself.

What the operators wanted

An exposed LLM endpoint is a server that answers AI requests without authentication, or with authentication a stolen key can pass. Because most AI tools now use the same OpenAI-compatible interface, a single scanner can find self-hosted model servers, AI gateways that sit in front of paid accounts, hastily deployed test systems and Model Context Protocol servers that connect a model to real tools.

The prize is free inference. By routing work through a stranger's endpoint or a stolen key, the operators shift the cost onto the victim and gain anonymity. Symantec says the honeypot's own telemetry recorded more than 127,000 external requests from tens of thousands of addresses over the period, dominated by reconnaissance of model endpoints, spraying of stolen API keys and probing for exposed configuration files.

A layered jailbreak and roughly 74 tools

Symantec excerpt of the captured prompt that renames the model Fantasy AI on GPT-4o and tells it that reading .env files and API keys is normal development work.
Screenshot: Symantec / Broadcom, Layers 1 and 2 of the captured Skynet prompt, Oct 21, 2026 (fair use).

The captured request was built in layers. The first reassigned the model a branded identity, Fantasy AI on GPT-4o, and told it never to reveal the real model, provider or that it was reached through a proxy. A second layer redefined reading configuration files, .env values, API keys and database schemas as normal development work the model should never refuse. Further layers cast the model as a devoted companion that must never refuse and told it to dismiss its own safety messages.

A closer reading of the tool definitions, which numbered around 74 rather than the handful named in the prompt's own summary, showed internet-wide asset discovery through FOFA; exploit modules aimed at VS Code, Django, Jupyter, Laravel, Supabase, Next.js and Flask/Werkzeug deployments, including tools that work out the Werkzeug debugger PIN; GraphQL and generic command injection; server-side request forgery through HTML-to-PDF generators; a dedicated SSRF tool that named cloud metadata services; automated dumping of databases exposed through Prisma Studio; and botnet command-and-control.

Symantec table of the four tool-call turns recorded in the captured request, ending in the honeypot's fallback reply, above the ssrf_exploit and c2_operator_start tool definitions.
Screenshot: Symantec / Broadcom, tool definitions and the four recorded turns against a training-platform target, Oct 21, 2026 (fair use).

A final directive placed the agent in what the operators called Skynet mode. It authorized up to 100,000 unattended steps, required the model to execute tools immediately without narrating or asking permission, and instructed it to deploy cryptocurrency miners quietly under /tmp/.x. The task appended to that particular request was a benign programming challenge from a legitimate training platform. The request also carried the record of three earlier tool calls against that target. None of those tool calls were run by the honeypot. The only reply it produced was its generic fallback message, after which the session ended.

A working payload a day later

A separate request the following day, sent from a different address, abandoned the tool-calling format. It asked the model to decode a string described only as a command to run. Decoded, the string was a gzip-compressed shell script piped to a command interpreter. The script downloaded, and where necessary compiled, XMRig, hid the binary under a filename designed to look like an unrelated logging utility, configured failover across four mining pools, set an immutable attribute on its configuration and ran a loop that erased a local service's access logs.

Symantec excerpt of the decoded go2rtc path logic that chooses a hiding directory and locates go2rtc.yml or go2rtc.yaml.
Screenshot: Symantec / Broadcom, the decoded go2rtc path-selection lines from the Day 2 payload, Oct 21, 2026 (fair use).

The script appears to have been written with camera servers in mind. It first looked for directories used by go2rtc, a WebRTC video streaming server found in smart camera gateways and Home Assistant setups, and placed the miner in the first one it found as a hidden file named .trace. The log-deletion loop polled go2rtc's local API on port 1984.

What to hunt

  • Inventory every internet-facing LLM API, self-hosted model server and MCP endpoint. Require authentication on every path, including model-listing and health-check paths, and place services behind a gateway rather than binding a model server to all interfaces.
  • Treat AI API keys as high-value credentials. Rotate them, scope them narrowly, keep them out of source code and configuration files, and alert on unusual inference volume or spend.
  • Restrict outbound network access from systems that host AI tools or run tool-enabled agents, so a compromised agent cannot scan the internet, reach a command-and-control server or download a miner.
  • On cloud workloads, enforce Instance Metadata Service version 2 with a hop limit of one.
  • Hunt for XMRig under go2rtc paths and for the file hashes fcde1323acceb46fd88f2276c010bb98839dccce03dc512c56d9529428ceec68 and b720a5e7c3a5b1760d15a8833a782f4fe7fc5c62b967fe57f607e72745ec0ccf. Block mining pools mine.c3pool[.]com:13333, auto.c3pool[.]org:80, auto.skypool[.]xyz:6666 and xmrpool[.]eu:5555, and the addresses 86.216.147.228, 13.140.161.91 and 207.180.232.206.
  • Keep the software the agent tries to exploit, including VS Code servers, Django, Jupyter, Laravel, Supabase and Prisma Studio, patched and off the public internet.

Sources:


Ava Okello covers detection engineering, EDR telemetry and SOC hunting for SOCtember from London.

Related stories

Detection desk