Symantec researchers describe an agent built for “LLMjacking”, where attackers take over someone else’s AI model. It scans the internet for reachable model servers, gateways and MCP servers, then uses the hijacked model to plan the attack. Instructions it was given include scanning targets, exploiting common web frameworks, stealing databases and credentials and installing a crypto miner, for up to 100,000 steps without human sign-off.
The researchers used a honeypot posing as an unprotected AI service. It logged over 127,000 outside requests. They warn that any organisation exposing AI services should assume they will be found and tested soon, and that unauthenticated endpoints are now actively hunted. The victim pays for the compute.
- Check self-hosted model servers and MCP endpoints for open access.
- Never leave an AI gateway without authentication.
Indian teams running internal AI prototypes on cloud servers should audit them. Security skills around AI infrastructure are a promising path for freshers.
Source: Help Net Security.