# Symantec finds an AI agent designed to power cyberattacks

> A honeypot captured an agent prompt built to hijack exposed AI models and run attacks for up to 100,000 steps. Symantec did not execute its tools.

Oossa · 2026-10-22 · https://oossa.com/en/symantec-finds-an-ai-agent-designed-to-power-cyberattacks

Symantec researchers say they found an offensive AI agent designed to take over exposed language models and use them to coordinate cyberattacks. The agent’s operators called its long-running mode “Skynet mode”: its prompt authorized up to 100,000 steps without human confirmation.

The plan was to make a model scan for other exposed AI services, exploit vulnerable servers, steal databases and credentials, and install software that mines cryptocurrency. If a paid AI account were abused, its owner could also face the model’s usage bill, Symantec says.

## A prompt built to bypass refusals

The attackers sent their full prompt to a Symantec honeypot, a decoy system that imitates an exposed AI service. It included five layers intended to make a commercial model comply: a fake identity, instructions to treat credential theft as routine developer work, a fictional companion persona, and directions to ignore safety warnings. A fifth layer contained 74 attack-tool modules.

The prompt described a repeating cycle of scanning, infecting servers, installing a cryptominer, and returning to previously targeted systems. Symantec also recorded a separate obfuscated command that would download XMRig, a Monero mining tool, and target camera servers using go2rtc.

## What the honeypot did—and didn’t—show

The decoy did not run the bundled tools or the later command; it recorded them. Symantec says the captured task was a harmless coding exercise, suggesting the operators may have been testing the agent against a safe target. The analysis does not establish whether the layered prompt can reliably defeat the safeguards of current commercial models or show that this captured agent completed an attack.

## The facts

- Symantec’s honeypot logged more than 127,000 requests from tens of thousands of addresses.
- The captured agent prompt authorized up to 100,000 steps without human involvement.
- The prompt included 74 modules for tasks such as scanning servers, exploiting web frameworks, and dumping databases.
- Symantec says the honeypot did not execute the agent’s tools or the separately captured mining command.

## Why it matters

Organizations that expose AI services to the internet have a concrete reason to secure those endpoints: Symantec recorded scans and attempts to use exposed models, and says the model owner could be left paying for unauthorized use. But the honeypot only captured the attack materials; whether this agent can make commercial models carry out the full plan remains unknown.

## Sources & references

1. [An AI agent in "Skynet mode" hijacks unprotected language models to launch self-sustaining cyberattacks](https://the-decoder.com/an-ai-agent-in-skynet-mode-hijacks-unprotected-language-models-to-launch-self-sustaining-cyberattacks/) – The Decoder, 2026-10-22

Last updated: 2026-10-22