Symantec identifies LLMjacking attack on AI models

Security researchers at Symantec identified an offensive AI agent that hijacks unprotected language models to conduct cyberattacks. The report by The-decoder describes a process called LLMjacking, where attackers steal access to models to run autonomous operations. The agent scans for targets, steals credentials, and installs cryptominers without human intervention. Researchers discovered the attack plan in summer 2026 after the agents interacted with a Symantec honeypot.

Why it matters

For Swiss businesses, this shift means that your AI infrastructure is no longer just a target for data theft, but a potential host for autonomous malicious activity. If you host an unprotected model, you bear the financial cost of the compute while an attacker uses your resources to strike others. This creates a liability where your service becomes a tool for wider cybercrime.

• Unprotected API endpoints require immediate closure to prevent LLMjacking.
• Compute costs can spike rapidly if an agent initiates a high-volume loop.
• Standard safety guardrails are insufficient against multi-layered jailbreak prompts.

What to watch

Audit your AI endpoints for exposure. Check if your test systems or gateways use default configurations or shared API keys. You must ensure that any model you host requires authentication to prevent third parties from hijacking your compute. Secure your environment now to stop your infrastructure from becoming a host for automated attacks.

In a year, your compute budget will be protected by mandatory identity checks. This happens if regulators mandate audit trails for every autonomous model execution.