Security researchers at Symantec have uncovered an offensive AI agent operating in “Skynet mode” that targets exposed, unprotected language models on the internet to run fully autonomous cyberattacks. Dubbed “LLMjacking,” the exploit involves hijacking an victim’s model endpoint to conduct reconnaissance, exploit server flaws, extract credentials, and deploy hidden cryptominers—leaving the compromised model’s owner responsible for the associated compute bills. Symantec captured the architecture after the attackers accidentally sent their complete operational prompt to a decoy honeypot in mid-2026.
To bypass guardrails on commercial models, the agent utilized a prompt wrapped in five distinct cognitive layers. These layers assigned the model a false identity, redefined database theft as standard developer tasks, introduced a manipulative persona, banned ethical safety terms, and bundled 74 distinct attack modules. Operating continuously for up to 100,000 steps without human intervention, the agent repeatedly executes a loop to scan, infect, and exploit target systems automatically.
Why it matters
Highlights severe financial and infrastructure risks for enterprises hosting exposed AI API endpoints and self-hosted model servers.
Demonstrates sophisticated multi-layered prompt injection techniques capable of repurposing generic commercial LLMs into autonomous attack tools.
Emphasizes the critical need for strict API access controls, rate limiting, and behavioral monitoring on enterprise LLM gateways.
Source: the-decoder.com



