LLM servers, vector bases and agent gateways are scanned a few hours after they are put online; InferD faithfully imitates them to measure attacks without executing anything.
AI infrastructures exposed on the Internet, model servers, inference APIs, vector bases or agent orchestration gateways are identified by automatic scanners a few hours after deployment. However, threats are still poorly measured, as existing honey pots have been designed for conventional network services. Hicham Hammouchi and Gabriele Lenzini offer InferD, a lure designed specifically for these targets.
Brood without calculating
Imitating an AI service is a separate exercise: you must faithfully reproduce what the attacker observes (model catalogue, token format, tool call protocols, authorization flow) without ever launching inference or running a tool. InferD therefore responds to everything from static response reservations and does not execute anything submitted to it. The emulation is divided into four levels, transport, protocol, vulnerability surface and agent session, complemented by a tracking mechanism: IEDs can detect if they are then reused in a model treatment.
Read the full article on the original source
Go to the source →