Symantec honeypot captures AI agent built for autonomous attacks
Original title: Symantec Uncovers AI Agent Hijacking Models for Autonomous Cyber-Attacks
Instructions covered up to 100,000 actions, including scanning, credential theft and Monero mining, and were not executed.
Symantec’s Threat Hunter Team says an AI honeypot captured the instruction set of an agent designed to hijack exposed LLM endpoints and run attacks without human approval. The agent sought self-hosted models, AI gateways and MCP servers—a practice researchers call LLMjacking—and included about 74 tools for reconnaissance, web exploitation, database theft, credential harvesting and botnet management. A separate request carried an obfuscated Monero-mining script. Symantec says the techniques are not new but their packaging into one self-directed agent lowers the skill needed for broad campaigns.
Unconfirmed: Symantec observed the agent’s instructions on a honeypot and did not confirm the tools ran against real victims.
- The honeypot logged over 127,000 external requests before the October 21 report.
- Instructions were captured in summer 2026; the honeypot did not run the tools.
- Layered prompts were meant to override commercial models’ safety controls.
- One request sought to install Monero-mining software, hide it and erase logs.
- Operators would shift model compute cost onto victims and gain anonymity.
Impact: Hijacked model compute used for scanning, theft and mining
- 1.Watch for exposed self-hosted LLMs, AI gateways and MCP servers.
- 2.Check endpoints for unauthorized use consistent with LLMjacking.
- ATT&CK
- T1595 Active ScanningT1190 Exploit Public-Facing ApplicationT1496 Resource HijackingT1070 Indicator Removal
- Vendors
- Broadcom
- Organizations
- Symantec Threat Hunter Team
From the original article
2 min read at infosecurity-magazine.com
Researchers at Symantec have uncovered an AI agent designed to operate a full cyber-attack without human intervention
By Kevin Poireault · infosecurity-magazine.com
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.