Uploaded November 2026 | Updated November 2026, 2 days ago
This talk examines how exposed AI infrastructure is becoming a new adversary playground, as attackers and internet-scale scanners and bruteforcing target LLM gateways, MCP servers, local inference APIs, and AI developer tooling faster than defenders have built threat models for them.
I ran a purpose-built AI honeypot that simulated 16 LLM and AI infrastructure personas across 16 ports, returning framework-authentic responses, headers, errors, and protocol behaviors. In one 48 hour window, the system captured 3,993 requests from 327 unique source IPs, including 155 MCP probes and 344 AI API key probes. What emerged was not generic internet noise, but a repeatable playbook against the emerging AI stack: LiteLLM model-registration abuse, MCP resource enumeration, framework-aware credential brute forcing, and coordinated scanning for exposed local inference services.
This talk examines how exposed AI infrastructure is becoming a new adversary playground, as attackers and internet-scale scanners and bruteforcing target LLM gateways, MCP servers, local inference APIs, and AI developer tooling faster than defenders have built threat models for them.
I ran a purpose-built AI honeypot that simulated 16 LLM and AI infrastructure personas across 16 ports, returning framework-authentic responses, headers, errors, and protocol behaviors. In one 48 hour window, the system captured 3,993 requests from 327 unique source IPs, including 155 MCP probes and 344 AI API key probes. What emerged was not generic internet noise, but a repeatable playbook against the emerging AI stack: LiteLLM model-registration abuse, MCP resource enumeration, framework-aware credential brute forcing, and coordinated scanning for exposed local inference services.










