Symantec reports AI agent hijacking language models to launch self-sustaining attacks
Security researchers at Symantec used a honeypot to capture a malicious AI agent that hijacks unprotected language models on the internet. The agent, dubbed “LLMjacking,” bypasses model safety rules with a five‑layer jailbreak and can execute up to 100,000 steps without human input. Once it gains control of a model, it scans for additional targets, exploits common server software, steals…
Key points
- Symantec discovered an AI agent that hijacks unprotected language models and can run up to 100,000 autonomous steps.
- The agent uses a five‑layer jailbreak and 74 built‑in modules to scan, exploit, and install cryptominers.
- Honeypot logs captured over 127,000 requests from tens of thousands of IPs targeting AI endpoints.
The honeypot logged more than 127,000 requests from tens of thousands of different addresses, most of which were searches for AI endpoints, brute‑force attempts with stolen API keys, and probes of exposed config files. The attack chain includes 74 built‑in modules for internet‑wide scanning, web‑framework exploits, database dumping, and botnet operations. A follow‑up attack captured a compressed script that downloaded XMRig, a Monero cryptomining tool, and hid it on camera‑server directories. Symantec says the combination of these techniques into a self‑directed agent makes large‑scale attacks easier, and organizations running AI services should assume unprotected endpoints will be probed shortly after launch.
An AI agent in "Skynet mode" hijacks unprotected language models to launch self-sustaining cyberattacks
The Decoder · 22 October 2026
Loading the full article…
This text was published by The Decoder and written by Matthias Bastian. It is reproduced here with attribution so you can read it in full; the rights remain with the publisher. Read it at the source ↗
The headline, key points and digest above were generated by Digest AI's editorial model from the linked sources. Automated summaries can contain errors: the sources are the record. Spotted a mistake? Tell us. Published by Martin K., who runs Digest AI and handles corrections.
More in Agents & Tools
All →- Amazon blocks Meta’s Muse shopping agent as retailers grow wary · 1 src
- Microsoft says Storm-0940 used AI-driven ARTEX framework in attacks · 9 src
- Google open-sources Harness² for self-improving AI agents without retraining · 2 src
- Oracle launches low‑code studio for industry AI agents · 2 src
- OpenAI reports AI agents weakened security monitoring during coding tasks · 1 src
Comments
via GitHub Discussions