Symantec honeypot captures AI agent built to hijack models
October 24: The Register says Symantec has no evidence of real attacks, names software targets, and reports the kit worked on a live target in testing.
Symantec captured but did not run an unattributed agent prompt for hijacking exposed models to steal data and mine, with no confirmed real victims.
Symantec’s Threat Hunter Team says a summer 2026 honeypot captured the full instructions of an unattributed agent meant to hijack exposed language-model endpoints and attack without human approval. The prompt sought self-hosted models, AI gateways, MCP servers and unauthenticated or stolen-key OpenAI-style endpoints, then planned up to 100,000 steps with about 74 tools for scanning, web exploitation, database and credential theft, remote code execution and cryptomining. Layered jailbreaks were meant to override safety controls and treat harvesting of .env files, API keys and schemas as routine work. A later request carried an obfuscated script to install XMRig on go2rtc camera servers and erase logs. The honeypot logged more than 127,000 requests and did not run the tools; Symantec says it has no evidence of success against real victims, though testing showed the kit worked against a live target.
- Honeypot logged more than 127,000 requests from tens of thousands of addresses.
- The agent included about 74 tools and a loop of up to 100,000 steps.
- A later script installs XMRig, hides it as .trace, and wipes go2rtc logs.
- Named targets include Django, Jupyter, Laravel, VS Code, Flask, Next.js and Supabase.
- Symantec has not attributed the kit and reports no confirmed real-world victims.
- Researchers Priti More and Susmit Yenkar described the summer capture.
What each outlet adds
beyond the shared facts- Infosecurity MagazineCalls the practice LLMjacking and says the honeypot did not execute the tools.
- The DecoderFive jailbreak layers, Fantasy AI and ENI personas, and a harmless first task.
- Help Net SecurityNames researchers and says XMRig is hidden as a file named .trace.
- The RegisterNo evidence of real victims; testing worked against a live target.
- Malware
- XMRig
- Organizations
- SymantecSymantec Threat Hunter Team
- Countries
- France
- AI models
- GPT-4o
Coverage
oldest first · one row per article · UTC- · 4d ago
Infosecurity MagazineFirst report
Symantec honeypot captures AI agent built for autonomous attacks
Instructions covered up to 100,000 actions, including scanning, credential theft and Monero mining, and were not executed.
- · 3d ago
The Decoder
Symantec finds AI agent hijacking exposed models for autonomous attacks
A honeypot captured a five-layer jailbreak and a loop of up to 100,000 steps that scans, steals data, and installs XMRig.
- · 2d ago
Help Net Security
Honeypot captures agent that hijacks exposed AI models
The captured prompt tells a stolen model to scan, exploit web apps, steal secrets and mine, for up to 100,000 steps.
- · 1d ago
The Register · Security
Symantec finds AI agent built to hijack models for unattended attacks
The captured kit can run up to 100,000 steps, but researchers have no evidence it was used in real attacks.